RATE OF CONVERGENCE IN CLT ON STRATIFIED GROUPS GYULA PAP ∗ Lajos Kossuth University, Debrecen, Hungary It is given a uniform estimate of the rate of convergence in CLT on stratified groups for variational distance when some pseudomoment is sufficiently small. Moreover it is given a nonuniform estimate on the system of homogeneous balls when some pseudomoment is finite. 1. Introduction. Let G be a stratified group of step s, that is a simply connected nilpotent Lie group whose Lie algebra G has a vector space decomposition G = ⊕ sj=1 Vj such that [Vi , Vj ] ⊂ Vi+j when i + j ≤ s and [Vi , Vj ] = 0 when i + j > s, and V1 generates G as an algebra. Let exp : G 7→ G be the exponential mapping (which is now a diffeomorphism). We equip G as well as G with the natural dilations by extending ◦ ◦ δ t (X) = tj X , t > 0, X ∈ Vj by linearity to G and putting δt (exp X) = exp(δ t X). The family (δt )t>0 is a continuous one-parameter semigroup of automorphisms of G. Let {X1 , . . . , Xm } be a basis of G adapted to the vector space decomposition G = that is {Xk : dk = j} is a basis for Vj where dk = j when Xk ∈ Vj . We denote by {ξ1 , . . . , ξm } the basis for the linear forms on G dual to the basis {X1 , . . . , Xm }, and we set ηj = ξj ◦ exp−1 . The functions {η1 , . . . , ηm } form a global coordinate system on G. ⊕sj=1 Vj , An element X ∈ G can be regarded as a (left-invariant) differential operator on G: for f ∈ C 1 we put f (x exp tX) − f (x) Xf (x) = lim . t→0 t The the following central limit theorem on G (cf. [4], [14]) is known . Let µ be a centered probability measure on G with finite second homogeneous moment. (That is, R R ηi (x)µ(dx) = 0 when di = 1 and |ηi (x)|2/di µ(dx) < ∞ for i = 1, . . . , m.) Then δ1/√n (µn ) → ν1 (1) where (νt )t>0 denotes the continuous convolution semigroup of Gaussian measures on G AMS 1980 subject classifications: Primary 60B15; secondary 60F05 Key words and phrases: Stratified groups, Gauss semigroups of measures on Lie groups, pseudomoments ∗ Research work supported by the Alexander von Humboldt Foundation and completed at the University of Tübingen, Federal Republic of Germany 1 whose infinitesimal generator is X 1X aij Xi Xj di =dj =1 di =2 2 R R where ai = ηi (x)µ(dx) for di = 2 and aij = ηi (x)ηj (x)µ(dx) for di = dj = 1. (For information on Gauss semigroups cf. [9], [16]). a i Xi + The aim of this paper is to examine the rate of convergence in (1). For the sake of simplicity it will be considered only the special case when ai = 0 for di = 2 and aij = δij for di = dj = 1. We obtain the optimal speed of convergence n−1/2 . We use the assumption of the finiteness of some pseudomoment and give estimates in terms of different kinds of moments and pseudomoments. We apply the method of composition. This has been used in this context for Euclidean spaces or Banach spaces; see for example [1], [2], [3], [11], [13] and [15]. For Heisenberg group see [12]; for that special group the results are a little bit better then the ones of the present paper. 2. Preliminaries. The letter c (respectively c(.) ) with or without indices will denote positive constants (respectively positive constants depending only on the quantities in parentheses); the same symbol may stand for different constants. ◦ ◦ It is known that if λ denotes the Lebesgue measure on G then λ = λ ◦ exp−1 is an invariant Haar measure on G. Let (νt )t>0 be the continuous convolution semigroup of Gaussian measures on G whose P infinitesimal generator is the sub-Laplacian (1/2) dk =1 Xk2 . Let R∗+ = {r ∈ R : r > 0}. The differential operator 1X ∂ Xk2 − dk =1 ∂t 2 is hypoelliptic on R∗+ × G by the Hörmander’s hypoellipticity criterion since {Xk : dk = 1} generates the whole Lie algebra G (cf. [10]; see also 6.3.7 in [9]). Thus the Gauss semigroup (νt )t>0 on G is absolutely continuous with respect to λ and there exists an infinitely differentiable function p > 0 on R∗+ × G such that pt (·) = p(t, ·) is a λ-density of the measure νt (cf. [16]). We shall write simply ν and p instead of ν1 and p1 , respectively. It is known that p(r 2 t, δr x) = r −Q p(t, x) (2) Pm for all x ∈ G, t > 0, r > 0, where Q = i=1 di is the homogeneous dimension of G. Taking into account that λ(δr (B)) = r Q λ(B) for all B ∈ B(G) we obtain νt = δ√t (ν1 ) for all t > 0. In particular, ν1n = νn = δ√n (ν1 ). Hence the measure ν is stable in the sense of Baldi and also stable with respect to the one-parameter semigroup (δ√t )t>0 of automorphisms of G in the sense of Hazod (see [7]). 2 A homogeneous norm on G is a function x → |x| from G to [0, ∞) satisfying (i) x → |x| is continuous on G and C ∞ on G \ {e}; (ii) |x| = 0 if and only if x = e; (iii) |δt x| = t|x| for t > 0, x ∈ G. Homogeneous norms always exist (cf. [5]). Moreover it is known that for a homogeneous norm | · | on G there is a constant c > 0 such that |xy| ≤ c(|x| + |y|) for all x, y ∈ G. A function f : G \ {0} 7→ R will be called homogeneous of degree α (α ∈ R) if f ◦ δr = rα f for r > 0. Clearly a homogeneous norm is a homogeneous function of degree 1. A linear differential operator D on G will be called homogeneous of degree α if D(f ◦ δr ) = r α (Df ) ◦ δr for any f ∈ C1 and r > 0. It can be observed that if D is homogeneous of degree α1 and f is homogeneous of degree α2 then Df is homogeneous of degree α2 − α1 . For example the differential operator Xk ∈ G is homogeneous of degree dk . LEMMA 1. Let | · | be a homogeneous norm on G and let f : G \ {0} 7→ R be a continuous function which is homogeneous of degree d ∈ R. Then there exists a constant c1 > 0 such that |f (x)| ≤ c1 |x|d for all x ∈ G \ {0}. If moreover f (x) 6= 0 for x ∈ G \ {0} then there exists a constant c2 > 0 such that |f (x)| ≥ c2 |x|d for all x ∈ G \ {0}. PROOF. Both sides of the desired inequalities are homogeneous of degree d, so it is suffices to assume that |x| = 1. The set {x ∈ G : |x| = 1} is compact (see 1.4 Lemma in [5]), and does not contain e ∈ G. Hence the assertion. # Let us define %(x) = Xm i=1 |ηi (x)|1/di for x ∈ G. The function % satisfies the assumptions of Lemma 1, thus for an arbitrary homogeneous norm | · | on G there exist c1 , c2 > 0 such that c1 %(x) ≤ |x| ≤ c2 %(x). (3) We adopt the following multiindex notation. Let Z+ = N ∪ {0}. If I = (i1 , . . . , im ) ∈ i1 I im im Zm and η I (x) = η1i1 (x) . . . ηm (x), x ∈ G. + is a multiindex, we set X = X1 . . . Xm Further, we set |I| = i1 + . . . + im and d(I) = i1 d1 + . . . + im dm . Thus |I| is the order of the differential operator X I , while d(I) is its degree of homogeneity (see [5]). A function P : G 7→ R will be called a polynomial if P ◦ exp is a polynomial on G. Every polynomial on G can be written uniquely as X P = aI η I I 3 where all but finitely many of the coefficients aI vanish. The homogeneous degree of a P polynomial P = I aI η I is max{d(I) : aI 6= 0}. Clearly η I is a homogeneous function of degree d(I), thus by Lemma 1 |η I (x)| ≤ cI |x|d(I) (4) for all x ∈ G with a suitable constant cI > 0. A polynomial P will be called homogeneous if it is homogeneous as a function. If P is a homogeneous polynomial which is homogeneous of degree d then it can be written uniquely as X P = aI η I d(I)=d and Lemma 1 implies that there is a constant cP > 0 such that |P (x)| ≤ cP |x|d for all x ∈ G. Let k ∈ N, f ∈ C k and x ∈ G. The left Taylor polynomial of f at x of homogeneous degree k is the unique polynomial P of homogeneous degree k such that X I P (e) = X I f (x) for I ∈ Zm + with d(I) ≤ k (see [5]). We shall use the following stratified Taylor inequality. THEOREM A. Let | · | be a homogeneous norm on G. For each k ∈ N there is a constant ck such that for all f ∈ C k and all x, y ∈ G, |f (xy) − Px(k) (y)| ≤ ck |y|k+1 supd(I)=k+1,|z|≤bk+1 |y| |X I f (xz)| where Pxk is the left Taylor polynomial of f at x of homogeneous degree k and b ≥ 1 is a constant (depending only on G and | · |) (cf. 1.44 Corollary in [5]). We shall apply also the following result: THEOREM B. Let | · | be a homogeneous norm on G. constants {CI , I ∈ Zm + } and C such that Then there exist positive |X I pt (x)| ≤ CI t−(d(I)+Q)/2 exp{−C|x|2 /t} for every t > 0, x ∈ G and I ∈ Zm + (cf. [8]). COROLLARY 1. For arbitrary k ∈ Z+ and I ∈ Zm + we have Z (1) |x|k |X I pt (x)|λ(dx) = ckI t(k−d(I))/2 Z ∂ I (2) |x|k pt (x) λ(dx) = ckI t(k−d(I))/2 ∂η Z (3) |x|k |X I pt (x)|λ(dx) ≤ ckI |x|>1 Z ∂ I (4) k pt (x) λ(dx) ≤ ckI . |x| ∂η |x|>1 4 (i) with some positive constants ckI , i = 1, 2, 3, 4. PROOF. From the homogeneity (2) we obtain X I pt (x) = t−(Q+d(I))/2 X I p1 (δ1/√t (x)). The integrability of the function x → |x|k |X I pt (x)| with respect to the measure λ follows from Theorem B and Corollary 1.17 in [5]. Substituting x = δ√t (y) and using λ(δ√t (dy)) = tQ/2 λ(dy) we obtain the first equation. It is known that ∂ ∂η I = X |J|≤|I|,d(J)≥d(I) PIJ X J (5) where PIJ is a homogeneous polynomial of degree d(J) − d(I) (cf. p.25 in [5]). Thus the inequality |PIJ (x)| ≤ cIJ |x|d(J)−d(I) , Theorem B and the homogeneity (2) imply the second equation. Applying again Theorem B and (2) we have Z Z k I (k−d(I))/2 k I |x| |X pt (x)|λ(dx) = t √ |y| |X p1 (y)|λ(dy) ≤ |y|>1/ t |x|>1 Z (3) (d(I)+k)/2 |X I p1 (y)|λ(dy) ≤ ckI . ≤ √ |y| |y|>1/ t One can prove the last inequality in the same way. # For the sake of simplicity we shall write ν and p instead of ν1 and p1 respectively, if this does not cause misunderstanding. 3. Taylor polynomial of homogeneous degree 2. LEMMA 2. Let f ∈ C 2 and x ∈ G. Then the left Taylor polynomial of f at x of homogeneous degree 2 is X 1X ηi (y)ηj (y)Xi Xj f (x). Px(2) (y) = f (x) + ηi (y)Xi f (x) + di =dj =1 di =1,2 2 (2) PROOF. The Taylor polynomial Px can be written uniquely as X 1X aij ηi (y)ηj (y). Px(2) (y) = a0 + ai ηi (y) + di =dj =1 di =1,2 2 (2) (2) We have X I Px (e) = X I f (x) for all I ∈ Zm Clearly Px (e) = a0 , + with d(I) ≤ 2. I thus a0 = f (x). If d(I) = 1, 2 then X = Xk with dk = 1, 2 or X I = Xk X` with 1 ≤ k ≤ ` ≤ m, dk = d` = 1. The Campbell-Hausdorff formula implies that X ηk (xy) = ηk (x) + ηk (y) + d(I)+d(J)=dk ,I6=0,J6=0 5 I J cIJ k η (x)η (y) (6) for all x, y ∈ G, k = 1, . . . , m (see p. 23 in [5]). Therefore Xk = X ∂ ∂ I[k] + c` η I d` ≥dk +1,d(I)=d` −dk ∂ηk ∂η` (7) where [k] is the multiindex with 1 in the k th place and zeros elsewhere (cf. 1.26 Proposition in [5]). Thus δki if di ≤ dk I[k] I (8) Xk ηi (y) = P c η (y) if d i ≥ dk + 1 d(I)=di −dk i Consequently Xk ηi (e) = δki and (Xk ηi ηj )(e) = 0. k ∈ {1, . . . , m} with dk = 1, 2, thus ak = Xk f (x). (2) We conclude Xk Px (e) = ak for Let now k, ` ∈ {1, . . . , m} with dk = d` = 1. If i ∈ {1, . . . , m} with di = 1 then we have X` ηi (y) = δ`i , thus Xk X` ηi (e) = 0. If di = 2 then (8) implies X` ηi (y) = P [j][`] ηj (y), thus dj =1 ci Xk X` ηi (y) = X [j][`] c Xk ηj (y) dj =1 i [k][`] = ci . Obviously ηk (x2 ) = 2ηk (x) since exp X exp X = exp(X +X) for all X ∈ G. Thus applying (6) with x = y we obtain X [k][`] dk =d` =1 ci ηk (x)η` (x) = 0 for i ∈ {1, . . . , m} with di = 2. Since x ∈ G can be arbitrary we conclude [k][`] ci [`][k] + ci = 0. (9) Applying (7) it can be shown that if di = dj = dk = d` = 1 then (Xk X` ηi ηj )(e) = ( 2 if i = j = k = ` 1 if i 6= j and i = k, j = ` or i = `, j = k 0 else. Let 1 ≤ k ≤ ` ≤ m with dk = d` = 1. We have Xk X` Px(2) (e) (2) = 2akk P [k][`] ak` + di =2 ai ci if k = ` if k = 6 ` thus Xk X` Px (e) = Xk X` f (x) implies akk = (1/2)Xk2 f (x) and ak` = (Xk X` − X [k][`] c Xi )f (x). di =2 i 6 (10) Now X` Xk is a right-invariant differential operator of order 2 and homogeneous degree 2, so it is a linear combination of the monomials X I such that d(I) ≤ 2 (see [5]). Consequently (2) we have again X` Xk Px (e) = X` Xk f (x). On the other hand 2akk if k = ` (2) [`][k] X` Xk Px (e) = a + P ac if k 6= ` k` thus ak` = (X` Xk − di =2 X i i [`][k] c Xi )f (x). di =2 i (11) Adding (10) and (11) and applying (9) we obtain ak` = (1/2)(Xk X` + X` Xk )f (x). Thus the assertion. # 4. Homogeneous moments and pseudomoments. Let µ be a probability measure on G. For k ∈ N consider the k th homogeneous moment of µ: Mk (µ) = m Z X i=1 |ηi (x)|k/di µ(dx). It should be noted that Mk (µ) < ∞ if and only if the measure µ has finite k th moment on G in the sense of [6], [14]. We use also the k th homogeneous pseudomoment of µ and ν defined by m Z X βk (µ, ν) = |ηi (x)|k/di |µ − ν|(dx) i=1 where |µ − ν| denotes the total variation of the signed measure µ − ν. The inequality (3) implies that for k ∈ N and for a homogeneous norm | · | on G there are constants (1) (2) ck , ck > 0 such that Z Z (2) (1) k |x|k µ(dx) |x| µ(dx) ≤ Mk (µ) ≤ ck ck (1) ck Z k |x| |µ − ν|(dx) ≤ βk (µ, ν) ≤ (2) ck Z |x|k |µ − ν|(dx). Let (νt )t>0 be the Gauss semigroup on G whose infinitesimal generator is the subP Laplacian (1/2) dk =1 Xk2 . We write simply ν instead of ν1 . R R LEMMA 3. We have ηi (x)ν(dx) = 0 for i = 1, ..., m and ηi (x)ηj (x)ν(dx) = δij for di = dj = 1. PROOF. Theorem B implies that all moments of ν exist. The measure ν is symmetric, R hence ηi (x)ν(dx) = 0 for all i = 1, ..., m. The central limit theorem (1) holds for the R measure µ = ν, thus ηi (x)ηj (x)ν(dx) = δij for di = dj = 1. # 7 The variational distance between ν and a probability measure µ on G can be estimated by the help of their pseudomoments in the following way. LEMMA 4. For all k ∈ N there exists a constant ck > 0 such that for every probability measure µ on G one has Q/(Q+k) |µ − ν|(G) ≤ ck βk (µ, ν). PROOF. We use the ideas of Lemma 1, p.12 in [15]. Suppose first that 0 < v = |µ − ν|(G) < 2. Let G = D + ∪ D− be a Hahn decomposition of G with respect to the signed measure µ − ν, i.e. D + and D − are disjoint Borel sets such that (µ − ν)(B) ≥ 0 if B ⊂ D + and (µ − ν)(B) ≤ 0 if B ⊂ D − . Define a measure ζ on B(G) by ζ(B) = ν(B ∩ D + ) + µ(B ∩ D − ). Then βk (µ, ν) ≥ ck Z k % (x)|µ − ν|(dx) ≥ ck Z %k (x)(ν − ζ)(dx). (12) Note that ζ(G) = 1 − v/2. Put Aa,r = {x ∈ G : %(a−1 x) < r} for a ∈ G, r > 0. Let r > 0 such that ν(Ae,r ) = v/2. Then ν(Ace,r ) = 1 − v/2 = ζ(Ae,r ) + ζ(Ace,r ) thus ζ(Ae,r ) = (ν − ζ)(Ace,r ). Consequently we have Z k Ae,r k k % (x)ζ(dx) ≤ r ζ(Ae,r ) = r (ν − and it follows that Z Z k % (x)(ν − ζ)(dx) ≥ ζ)(Ace,r ) k Ae,r % (x)(ν − ζ)(dx) + Now (12) and (13) imply βk (µ, ν) ≥ ck Z Z ≤ Z Ace,r k %k (x)(ν − ζ)(dx) % (x)ζ(dx) = Ae,r Z %k (x)ν(dx). (13) Ae,r %k (x)ν(dx). Ae,r Let h ∈ (0, r) such that ν(Ae,h ) = ν(Ae,r \ Ae,h ) = v/4. Using that the density function of ν is bounded we have Z Z Q λ(dy) ≤ c, hQ . ν(Ae,h ) ≤ cλ(Ae,h ) = c λ(dx) = ch %(y)<1 %(x)<h 8 Consequently h ≥ c,, v 1/Q and Z βk (µ, ν) ≥ ck %k (x)ν(dx) ≥ ck hk ν(Ae,r \ Ae,h ) ≥ ck hk v/4 ≥ c,k v (Q+k)/Q . Ae,r \Ae,h Q/(Q+k) From this it follows v = |µ − ν|(G) ≤ c,,k βk (µ, ν). If |µ − ν|(G) = 0 then the statement is obvious. If |µ − ν|(G) = 2 then µ and ν are orthogonal, and we have Z k % (x)|µ − ν|(dx) Hence the assertion. Q/(Q+k) ≥ Z k % (x)ν(dx) Q/(Q+k) = ck = ck |µ − ν|(G)/2. # The variational distance between ν and the normalized convolution power δ 1/√n (µn ) can be estimated as follows. LEMMA 5. For all n ∈ N there exists a constant cn > 0 such that if µ is a probability R R measure on G with ηi (x)µ(dx) = 0 for di = 1, 2 and ηi (x)ηj (x)µ(dx) = δij for di = dj = 1 then for all n ∈ N |δ1/√n (µn ) − ν|(G) ≤ (|µ − ν|(G))n + cn X3s j=3 βj (µ, ν). (14) PROOF. We use again the ideas of Lemma 1, p.12 in [15]. Let us first observe that |(µ − ν)n (B)| ≤ 1 (|µ − ν|(G))n 2 (15) for all n ∈ N and B ∈ B(G). Indeed this is true for n = 1 and for n ≥ 2 using induction on n we have Z n n−1 −1 |(µ − ν) (B)| = (µ − ν) (x B)(µ − ν)(dx) ≤ Z 1 1 n−1 |µ − ν|(dx) = (|µ − ν|(G))n . ≤ (|µ − ν|(G)) 2 2 Let us now estimate |ν ∗ (µ − ν)(B)| for B ∈ B(G). We have ZZ ZZ ν ∗ (µ − ν)(B) = χB (yx)p(y)λ(dy)(µ − ν)(dx) = χB (u)p(ux−1 )λ(du)(µ − ν)(dx). Now we shall use Taylor inequality for the function Z f (x) = χB (u)p(ux−1 )λ(du), 9 x∈G at e ∈ G. Put gu (x) = p(ux−1 ) for x, u ∈ G. First we show that f ∈ C ∞ and Z I X f (x) = χB (u)X I gu (x)λ(du) for all x ∈ G and I ∈ Zm + . We have I X gu (x) = X |J|≤|I|,d(J)≥d(I) PIJ (x) ∂ ∂η J (16) gu (x) where PIJ is a homogeneous polynomial of degree d(J) − d(I) (cf. [5, p.25]). Moreover, from (6) J K X ∂ ∂ gu (x) = p(ux−1 ) QJK (x, u) |K|≤|J|,d(K)≥d(J) ∂η ∂η where QJK is a polynomial on G × G which is jointly homogeneous of degree d(K) − d(J) (that is, QJK (δr (x), δr (u)) = r d(K)−d(J) QJK (x, u)). By (4), Corollary 1 and by the inequality |vx| ≤ c(|v| + |x|) J Z ∂ gu (x) λ(du) χB (u) ∂η K Z X ∂ −1 ≤ p(ux ) λ(du) QJK (x, u) |K|≤|J|,d(K)≥d(J) ∂η K Z X (17) ∂ ≤ p(v) λ(dv) QJK (x, vx) |K|≤|J|,d(K)≥d(J) ∂η Z ∂ K X ≤ cJ (|v| + |x|)d(K)−d(J) p(v) λ(dv) |K|≤|J|,d(K)≥d(J) ∂η ≤ c(J, R) if |x| ≤ R. Thus ∂ ∂η J f (x) = Z χB (u) ∂ ∂η J gu (x)λ(du) which implies (16). Denote P (2) the left Taylor polynomial of f at e ∈ G of homogeneous degree 2. The moment conditions, Lemma 2 and Lemma 3 imply Z P (2) (x)(µ − ν)(dx) = 0. By Theorem A we have |f (x) − P (2) (x)| ≤ |x|3 supd(I)=3,|z|≤b3 |x| |X I f (z)| 10 for all x ∈ G. For |z| ≤ b3 |x| by (16) and (17) K Z ∂ I p(v) λ(dv) |X f (z)| ≤ |PIJ (z)| QJK (z, vz) |K|≤|J|≤|I| ∂η d(K)≥d(J)≥d(I) X ≤ c(I) |x|d(J)−d(I) 1 + |x|d(K)−d(J) . |K|≤|J|≤|I| X d(K)≥d(J)≥d(I) Consequently Z Z (2) |ν ∗ (µ − ν)(B)| = f (x)(µ − ν)(dx) = (f (x) − P (x))(µ − ν)(dx) Z Z 3 ≤ c |x| supd(I)=3,|z|≤b3 |x| |X I f (z)||µ − ν|(dx) Z X X |x|d(J)−d(I) + |x|d(K)−d(I) |µ − ν|(dx) ≤ c |x|3 |K|≤|J|≤|I| d(I)=3 ≤c X 3≤j≤3s Z d(K)≥d(J)≥d(I) |x|j |µ − ν|(dx) since d(I) = 3 and |K| ≤ |I| imply d(K) ≤ s|K| ≤ s|I| ≤ sd(I) = 3s. Thus we conclude |ν ∗ (µ − ν)(B)| ≤ c X3s βj (µ, ν). (18) |(µ − ν) ∗ ν(B)| ≤ c X3s βj (µ, ν). (19) Similarly j=3 j=3 Now we prove (14). It is true for n = 1. Suppose that it is true for 1, 2, . . . , n − 1. Then for all Borel sets B ∈ B(G) we have |(δ1/√n (µn ) − ν)(B)| = |(µn − ν n )(δ√n (B))| ≤ |(µ − ν)n (δ√n (B))| + S1 + S2 where n−1 X (µ − ν)k ∗ ν ∗ µn−k−1 (δ√n (B)) S1 = k=1 n−1 X S2 = ν k ∗ (µ − ν) ∗ µn−k−1 (δ√n (B)) . k=1 From (15), (18), (19) and (20) we obtain (14). # 5. Rate of convergence in CLT for variational distance. 11 (20) TEOREM 1. There exist two constants C1 , C2 > 0 with the following property: if µ is R R a probability measure on G such that ηi (x)µ(dx) = 0 when di = 1, 2, ηi (x)ηj (x)µ(dx) = δij when di = dj = 1 and β3s (µ, ν) ≤ C1 then for all n ≥ 4 one has ∆n (µ, ν) = X3s−3 sup δ1/√n (µn )(B) − ν(B) ≤ C2 β3+j (µ, ν)n−(1+j)/2 . j=0 B∈B(G) (21) REMARK. Lemma 4 and Lemma 5 give an estimate for ∆n (µ, ν) when n = 1, 2, 3. PROOF. We use induction on n and the method of composition, and follow the ideas of the proof of Lemma 2 in [2]. By the assumption β3s ≤ C1 , Lemma 4 and Lemma 5 X3s 1 1 √ 4 4 (|µ − ν|(G)) + c βj ≤ ∆4 (µ, ν) ≤ δ1/ 4 (µ ) − ν (G) ≤ j=3 2 2 X3s X3s 3(Q−s)/(Q+3s) 4Q/(Q+3s) ≤ c β3s + βj ≤ c 1 + C 1 βj j=3 j=3 since Q ≥ m ≥ s. We conclude that (21) is true for n = 4 if we choose C2 ≥ 4(1+3s)/2 c(1 + 3(Q−s)/(Q+3s) C1 ). Now we assume that if β3s (µ, ν) ≤ C1 then X3s−3 √ k β3+j (µ, ν)k −(1+j)/2 = Sk sup δ1/ k (µ )(B) − ν(B) ≤ C2 j=0 B∈B(G) (22) for k = 2, 3, . . . , n − 1 and B ∈ B(G). Consider the decomposition δ1/√n (µn ) where −ν = n X (γk + γ̃k ) (23) k=1 γk = (δ1/√n ν)k−1 ∗ (δ1/√n µ − δ1/√n ν) ∗ ((δ1/√n µ)n−k − (δ1/√n ν)n−k ) γ̃k = (δ1/√n ν)k−1 ∗ (δ1/√n µ − δ1/√n ν) ∗ (δ1/√n ν)n−k . First we give an estimate for γ1 (B), B ∈ B(G) : γ1 (B) = Z (δ1/√n µn−1 − δ1/√n ν n−1 )(x−1 B)(δ1/√n µ − δ1/√n ν)(dx). Using the inductive hypothesis (22) for n − 1 we have |(δ1/√n µn−1 − δ1/√n ν n−1 )(x−1 B)| = |(δ1/√n−1 µn−1 − ν)(δ√n/(n−1) (x−1 B))| ≤ Sn−1 12 thus using Sn−1 ≤ 2(3s−2)/2 Sn we obtain by Lemma 4 Q/(Q+3s) |γ1 (B)| ≤ Sn−1 |δ1/√n µ − δ1/√n ν|(G) ≤ cSn |µ − ν|(G) ≤ c, β3s Sn . (24) Now we give an estimate of γk (B) for 2 ≤ k ≤ [n/2] : γk (B) = ZZ (δ1/√n µn−k − δ1/√n ν n−k )(u−1 B)pt (ux−1 )λ(du)(δ1/√n µ − δ1/√n ν)(dx) where t = (k − 1)/n. We use now the Taylor inequality for the function f : G 7→ R, f (x) = Z (δ1/√n µn−k − δ1/√n ν n−k )(u−1 B)pt (ux−1 )λ(du) as in the proof of Lemma 5. Put gut (x) = pt (ux−1 ) for x, u ∈ G and t > 0, and l(u) = δ1/√n (µn−k − ν n−k )(u−1 B) for u ∈ G. From the inductive hypothesis (22) for n − k we have |l(u)| ≤ Sn−k ≤ cSn . (25) As in the proof of Lemma 5 one can show that f ∈ C ∞ and I X f (x) = Z l(u)X I gut (x)λ(du). Denote P (2) the left Taylor polynomial of f at e ∈ G of homogeneous degree 2. By the moment conditions we have again Z P (2) (x)δ1/√n (µ − ν)(dx) = 0. By Theorem A we have |f (x) − P (2) (x)| ≤ |x|3 supd(I)=3,|z|≤b3 |x| |X I f (z)| (26) for all x ∈ G. By (25) we have as in the proof of Lemma 5 for |z| ≤ b3 |x| |X I f (z)| ≤ ≤ c I Sn ≤ c I Sn X X |K|≤|J|≤|I| d(K)≥d(J)≥d(I) |K|≤|J|≤|I| d(K)≥d(J)≥d(I) ∂ K d(J)−d(I) d(K)−d(J) d(K)−d(J) pt (v) λ(dv) |x| |x| + |v| ∂η |x|d(K)−d(I) t−d(K)/2 + |x|d(J)−d(I) t−d(J)/2 Z 13 taking into account Corollary 1. Thus by (26) Z Z |γk (B)| = f (x)δ1/√n (µ − ν)(dx) = (f (x) − P (2) )δ1/√n (µ − ν)(dx) ≤ Z X3s −j/2 ≤ cSn |x|j |δ1/√n (µ − ν)|(dx) ≤ t j=3 ≤ cSn Consequently X3s j=3 ((k − 1)/n)−j/2 βj n−j/2 = cSn X[n/2] k=2 |γk (B)| ≤ cSn X3(s−1) X3s−3 j=0 j=0 β3+j (k − 1)−(3+j)/2 . β3+j . (27) If [n/2] < k ≤ n then instead of (25) we use the simple estimate |(δ1/√n µn−k − δ1/√n ν n−k )(u−1 B)| ≤ 1 (since both δ1/√n µn−k (u−1 B) and δ1/√n ν n−k )(u−1 B) are in [0, 1]), and t = (k − 1)/n ≥ 1/2 implies X3(s−1) |γk (B)| ≤ c β3+j n−(3+j)/2 . j=0 The same estimate is valid for γ̃k . Thus Xn k=[n/2]+1 Xn k=1 |γk (B)| ≤ c |γ̃k (B)| ≤ c j/k X3(s−1) j=0 X3(s−1) j=0 β3+j n−(1+j)/2 ≤ cC2−1 Sn (28) β3+j n−(1+j)/2 ≤ cC2−1 Sn . (29) Hölder inequality implies βj ≤ βk for 1 ≤ j ≤ k. Collecting the estimates (24) and (27)—(29) we obtain from (23) that X3s−3 Q/(Q+3s) −1 n |(δ1/√n µ − ν)(B)| ≤ c β3s + β3+j + C2 Sn j=0 X3s−3 (3+j)/3s Q/(Q+3s) −1 ≤ c C1 + C1 + C2 Sn j=0 ≤ Sn taking C1 sufficiently small and C2 sufficiently large. The proof is complete. # COROLLARY 2. There exist two constants C1 , C2 > 0 such that for every probability measure µ on G satisfying the assumptions of Theorem 1 and for all n ≥ 4 ∆n (µ, ν) ≤ C2 (β3 (µ, ν)n−1/2 + β3s (µ, ν)n−(3s−2)/2 ) ∆n (µ, ν) ≤ C2 (m3 (µ)n−1/2 + m3s (µ)n−(3s−2)/2 ). 14 6. Homogeneous balls. Let | · | be a homogeneous norm on G. For a ∈ G and r > 0 we define B(a, r) = {x ∈ G : |a−1 x| < r} and we call B(a, r) the homogeneous ball of radius r about a. We observe that δt (bB(a, r)) = B(δt (ba), rt) for all a, b ∈ G and r, t > 0. In other words, the system {B(a, r) : a ∈ G, r > 0} of homogeneous balls is closed with respect to the dilations (δt )t>0 and translation from the left. Let us consider now the homogeneous norm on G defined by |x| = m X i=1 |ηi (x)| d/di !1/d (30) for x ∈ G, where d ∈ N and d/di are even numbers for i = 1, . . . , m. We shall show that the system of homogeneous balls with respect to the above homogeneous norm has two important properties. LEMMA 6. Let | · | be the homogeneous norm on G defined by (30). Then there exists a constant c > 0 such that ν(B(a, r + ε) \ B(a, r)) ≤ cε(1 + |a|)s−1 for all a ∈ G, r > 0 and ε > 0. PROOF. To avoid complicated notation we write xi = ηi (x) for x ∈ G, i = 1, . . . , m in the proof. For r > 0, i = 1, . . . , m consider the sets n X Ai = x ∈ G : 1≤j≤m,j6=i |xj |d/dj ≤ c0 rd o where (m − 1)/m < c0 < 1 is a fixed constant. It is easy to notice that for r > ((c0 m/(m − 1))1/d − 1)−1 ε = c1 ε we have m \ i=1 thus Aci ⊂ (B(e, r + ε))c B(e, r + ε) \ B(e, r) ⊂ m [ i=1 (Ai ∩ (B(e, r + ε) \ B(e, r))). 15 Applying the translation invariance of the homogeneous balls we can estimate in the following way: Z ν(B(a, r + ε) \ B(a, r)) = = where Ii = Z Z p(x)λ(dx) = Z B(a,r+ε)\B(a,r) B(e,r+ε)\B(e,r) p(ay)λ(dy) ≤ m X Ii i=1 p(ay)λ(dy) Ai ∩(B(e,r+ε)\B(e,r)) = P j6=i |yj |d/dj ≤c0 r d Z p(ay)λ(dy). {yi :r≤|y|<r+ε} Let us fix y1 , . . . , yi−1 , yi+1 , . . . , ym ∈ R. We remark that {yi : r ≤ |y| < r + ε} = (−bi , −ai ] ∪ [ai , bi ) where ai (y) = (r d − X |yj |d/dj )di /d X bi (y) = ((r + ε)d − |yj |d/dj )di /d . j6=i j6=i For y ∈ Ai ∩ (B(e, r + ε) \ B(e, r)) we have (taking also into account r ≥ c1 ε): bi (y) − ai (y) ≤ ε supr≤z≤r+ε di z d−1 (z d − X j6=i |yj |d/dj )di /d−1 ≤ εdi (r + ε)d−1 (1 − c0 )di /d−1 rdi −d ≤ cεr di −1 . Applying Theorem B we obtain Z bi ai p(ay)dyi ≤ cεr di −1 supai ≤yi ≤bi exp{−C|ay|2 }. The supremum can be estimated in the following way: Z 2 supai ≤yi ≤bi exp{−C|ay| } ≤ |∂i exp{−C|ay|2 }|dyi . yi ≥ai By formula (6) we obtain ∂i |ay| = |ay|1−d We have X dj ≥di d/dj −1 d−1 j (ηj (ay)) ∂i = X dk ≥di 16 X d(I)+d(J)=dj Pik Xk I J cIJ j η (a)∂i η (y). where Pik is a homogeneous polynomial of degree dk − di (cf. p.25 in [5]). Since Xk η J is homogeneous of degree d(J) − dk Lemma 1 implies X X |∂i η J (y)| ≤ |Pik (y)||Xk η J (y)| ≤ c |y|dk −di |y|d(J)−dk ≤ c, |y|d(J)−di . dk ≥di dk ≥di If d(J) < di then clearly ∂i η J (y) = 0 for all y ∈ G. Using also (4) we obtain X X ∂i |ay| ≤ c |ay|1−dj |a|d(I) |y|d(J)−di . dj ≥di d(I)+d(J)=dj Collecting the above estimates we have Z Z di −1 Ii ≤ cεr P ≤ cε = cε Z Z j6=i |yj | d/dj ≤c0 rd |y|di −1 exp{−C|ay|2 } exp{−C|x|2 } ≤ c, ε(1 + |a|s−1 ) X |yi |≥ai X dj ≥di |∂i exp{−C|ay|2 }|λ(dy) dj ≥di |x|2−dj |ay|2−dj X X d(I)+d(J)=dj d(I)+d(J)=dj |a|d(I) |y|d(J)−di λ(dy) |a|d(I) |a−1 x|d(J)−1 λ(dy) where the following integrability condition was used: if f is a measurable function on G such that |f (x)| = O(|x|α−Q ) where α > 0 then f is integrable near 0 (cf. p.15 in [5]). In case r < c1 ε we can simply use the boundedness of the density function p: ν(B(a, r + ε) \ B(a, r)) ≤ cλ(B(a, r + ε) \ B(a, r)) = c, ((r + ε)Q − rQ ) ≤ c, εQ ≤ c, ε if ε ≤ 1. In case ε > 1 the proposition is trivial, since we can choose c = 1. # Let us fix an infinitely differentiable function ψ : R 7→ [0, 1] with ψ(z) = 1 for z ≤ 0 and ψ(z) = 0 for z ≥ 1. Let r, ε > 0. Put ( d |x| −(r−ε)d if r > ε (31) Ψ(x) = ψ rd −(r−ε)d 0 if r ≤ ε for x ∈ G. It is easy to show that we have χB(e, r − ε) ≤ Ψ ≤ χB(e, r) where χA denotes the indicator function of the set A. LEMMA 7. Let | · | be the homogeneous norm on G defined by (30). Let r, ε > 0. Let Ψ be defined by (31). Then for all I ∈ Zm + there exists a constant cI > 0 such that ∂ I Ψ(x) ≤ cI ε−|I| r|I|−d(I) ∂η 17 for all x ∈ G. PROOF. Let I = (i1 , . . . , im ) ∈ Zm + . We can assume r > ε. Since Ψ(x) = 0 for x ∈ G with |x| ≤ r − ε or |x| ≥ r it can be assumed r − ε ≤ |x| ≤ r. We have ∂ I X Ym |ηk (x)|jk d/dk −ik Ψ(x) ≤ c(I) . k=1 (r d − (r − ε)d )jk il dl /d≤jl ≤il , 1≤l≤m ∂η For jk ≤ ik we have 1 − (1 − ε/r)d ≥ ε/r ≤ (ε/r)ik /jk , thus (r d − (r − ε)d )jk ≥ εik rjk d−ik . Applying also the inequality |ηk (x)| ≤ |x|dk ≤ rdk we obtain the assertion. # 7. Rate of convergence in CLT on homogeneous balls. following ‘smoothing inequality’ (cf. [11], [13, Chapter 5, Lemma 1.1]). We shall use the LEMMA 8. Let µ1 and µ2 be probability measures on a measurable space (Ω, F). Let A, A1 , A2 ∈ F such that A1 ⊂ A ⊂ A2 and let ϕ1 and ϕ2 be measurable functions on (Ω, F) with χA ≤ ϕ1 ≤ χA ≤ ϕ2 ≤ χA . Then 1 2 Z |µ1 (A) − µ2 (A)| ≤ max ϕi (x)(µ1 − µ2 )(dx) + min µi (A2 \ A1 ). i=1,2 i=1,2 Ω TEOREM 2. Let | · | be a homogeneous norm on G with the following two properties: (i) There exists a constant c > 0 such that ν(B(a, r + ε) \ B(a, r)) ≤ cε(1 + |a|)s−1 (32) for all a ∈ G, r > 0 and ε > 0; (ii) for r > ε > 0 there exists a function Ψr,ε : G 7→ [0, 1] such that χB(e, r − ε) (x) ≤ Ψr,ε (x) ≤ χB(e, r) (x) for all x ∈ G, and ∂ I Ψr,ε (x) ≤ cI ε−|I| r|I|−d(I) ∂η for all x ∈ G and for all I ∈ Zm + whith d(I) ≤ 3s. (33) (34) Then there exist a constant C > 0 with the following property: if µ is a probability R R measure on G such that ηi (x)µ(dx) = 0 when di = 1, 2, ηi (x)ηj (x)µ(dx) = δij when R 3s/di di = dj = 1, |ηi (x)| µ(dx) < ∞ when i = 1, . . . , m, then for all n ≥ 4, a ∈ G and r > 0 one has δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ Cκ(a, r)(β3 (µ, ν) + β3s (µ, ν))n−1/2 (35) 18 where κ(a, r) = (1 + |a|)s−1 (1 + (1 + |a|)/r})3(s−1) . REMARK. Lemma 6 and Lemma 7 show that there exist always homogeneous norms with properties (i) and (ii) on arbitrary stratified groups. PROOF. We use again induction on n and the method of composition and follow the ideas of the proof of Theorem 1 in [2]. If β3s > 1 then using Lemma 4 we have |(µ − ν)(B)| ≤ 1 Q/(Q+3s) |µ − ν|(G) ≤ cβ3s ≤ cβ3s 2 for all B ∈ B(G), thus (35) is true for n = 1 (choosing C ≥ c), and we begin the induction with n = 1. If β3s ≤ 1 then as at the beginning of the proof of Theorem 1 X3s 4Q/(Q+3s) |(δ1/√4 (µ4 ) − ν)(B)| ≤ c(β3s + βj ) j=3 X3s βj ≤ c,, (β3 + β3s ) ≤ c, j=3 for all B ∈ B(G), since βj ≤ cj (β3 + β3s ) for 3 ≤ j ≤ 3s. Thus (35) is true for n = 4 (choosing C ≥ 2c), and we begin the induction with n = 4. Now we assume that √ k δ1/ k (µ )(B(a, r)) − ν(B(a, r)) ≤ Cκ(a, r)(β3 (µ, ν) + β3s (µ, ν))k −1/2 = Sk (36) for a ∈ G, r > 0 and k ≤ n − 1. Let now a ∈ G and r, ε > 0. Denote by Ψr,ε the function on G with properties (33) and (34) when r > ε and put Ψr,ε (x) ≡ 0 for x ∈ G when r ≤ ε. Put Ψ1 = Ψr,ε and Ψ2 = Ψr+ε,ε . Put ϕi (x) = Ψi (a−1 x) for x ∈ G, i = 1, 2. Then χB(a, r − ε) ≤ ϕ1 ≤ χB(a, r) ≤ ϕ2 ≤ χB(a, r + ε) . (37) Applying Lemma 8 and the assumption (32) we obtain δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ Z n √ ≤ max ϕi (x)(δ1/ n µ − ν)(dx) + cε(1 + |a|)s−1 . i=1,2 We shall write from now on ϕ and Ψ instead of ϕi and Ψi . 19 (38) We use again the decomposition (22). First we suppose that ε < min{1, r/2} and we R shall give an estimate for ϕ(x)(γk + γ̃k )(dx) if 1 ≤ k ≤ [n/2]. We have Z ZZZ ϕ(x)(γk + γ̃k )(dx) = ϕ(yuz)δ1/√n ν k−1 (dy)δ1/√n (µ − ν)(du)δ1/√n µn−k (dz). Now we use the Taylor inequality for the function ZZ f (u) = ϕ(yuz)δ1/√n ν k−1 (dy)δ1/√n µn−k (dz), u∈G at e ∈ G. Put gyz (u) = ϕ(yuz) = Ψ(a−1 yuz) for y, u, z ∈ G. First we show that f ∈ C 3 and ZZ I X f (u) = X I gyz (u)δ1/√n ν k−1 (dy)δ1/√n µn−k (dz) (39) for all u ∈ G and I ∈ Zm with d(I) ≤ 3. Since the map (y, u, z) → ηk (yuz) + is a polynomial on G × G × G which is jointly homogeneous of degree dk (that is, ηk (δr (y)δr (u)δr (z)) = r dk ηk (yuz)) we have X ηk (a−1 yuz) = η I (a−1 y)η J (u)η K (z) cIJK k d(I)+d(J)+d(K)=dk for some constants cIJK . Therefore k ∂ ∂η J gyz (u) = X |K|≤|J|,d(K)≥d(J) PJK (a −1 y, u, z) ∂ ∂η K Ψ(a−1 yuz) (40) where PJK is a polynomial on G×G×G which is jointly homogeneous of degree d(K)−d(J). Moreover, J X ∂ I QIJ (u) X gyz (u) = gyz (u) (41) |J|≤|I|,d(J)≥d(I) ∂η where QIJ is a homogeneous polynomial on G of degree d(J) − d(I). Let us consider the case gyz (u) = ϕ2 (yuz) = Ψ2 (a−1 yuz). The case gyz (u) = ϕ1 (yuz) = Ψ1 (a−1 yuz) can be handled similarly. The assumption (33) implies that Ψ2 (a−1 yuz) = 0 if |a−1 yuz| > r + ε. If |a−1 yuz| ≤ r + ε we have |z| ≤ c(|u−1 y −1 a| + |a−1 yuz|) ≤ c(|u| + |a−1 y| + r + ε). (42) For K ∈ Zm + with |K| ≤ |I| ≤ d(I) ≤ 3 we have d(K) ≤ 3s, and assumption (34) implies ∂ K −1 Ψ2 (a yuz) ≤ cK ε−|K| (r + ε)|K|−d(K) . (43) ∂η 20 By (40)—(43) and applying Theorem B we obtain ZZ |X I gyz (u)|δ1/√n ν k−1 (dy)δ1/√n µn−k (dz) ≤ c(I, ε, r, R) for u ∈ G with |u| ≤ R, which implies (39). Denote P (2) the left Taylor polynomial of f at e ∈ G of homogeneous degree 2. By the moment conditions we have again Z P (2) (x)δ1/√n (µ − ν)(dx) = 0. Thus by Lemma 2 we obtain Z where ϕ(x)(γk + γ̃k )(dx) = I0 − I1 − I2 + I3 Z (f (u) − f (e))δ1/√n (µ − ν)(du) |u|>1 Z X I1 = ηi (u)Xi f (e)δ1/√n (µ − ν)(du) di =1,2 |u|>1 Z X I2 = ηi (u)ηj (u)Xi Xj f (e)δ1/√n (µ − ν)(du) di =dj =1 |u|>1 Z I3 = (f (u) − P (2) (u))δ1/√n (µ − ν)(du). I0 = |u|≤1 Using |ϕ(yuz) − ϕ(yz)| ≤ 1 for y, u, z ∈ G we obtain |f (u) − f (e)| ≤ 1 for u ∈ G thus |I0 | ≤ |δ1/√n (µ − ν)|({u ∈ G : |u| > 1}) ≤ cβ3 n−3/2 . Now we shall estimate I X gyz (e) = ∂ ∂η I gyz (e). We consider again only the case gyz (u) = ϕ2 (yuz) = Ψ2 (a−1 yuz). We have ∂ ∂η K Ψ2 (a−1 yz) = 0 if |a−1 yz| ≤ r or |a−1 yz| ≥ r + ε, that is, if z ∈ / B(y −1 a, r + ε) \ B(y −1 a, r). Thus δ1/√n µ n−k ( z∈G: ∂ ∂η I gyz (e) 6= 0 )! ≤ δ1/√n µn−k (B(y −1 a, r + ε) \ B(y −1 a, r)). 21 Using ε < r/2, k ≤ [n/2] and the inductive hypothesis (36) for n − k we have |δ1/√n (µn−k − ν n−k )(B(y −1 a, r + ε))| = |δ1/√n−k (µn−k − ν n−k )(δ√n/(n−k) (B(y −1 a, r + ε)))| p ≤ Cκ δ√n/(n−k) (y −1 a), (r + ε) n/(n − k) (β3 + β3s )(n − k)−1/2 ≤ cC(1 + |a| + |y|)s−1 (1 + (1 + |a| + |y|)/r)3(s−1) (β3 + β3s )n−1/2 . In the same way we obtain the same estimate for |δ1/√n (µn−k − ν n−k )(B(y −1 a, r))|. From the assumption (32) follows ν δ√n/(n−k) B(y −1 a, r + ε) \ B(y −1 a, r) ≤ cε(1 + |a| + |y|)s−1 . Consequently for fixed a, y ∈ G we have δ1/√n µn−k z ∈ G : X I gyz (e) 6= 0 ≤ c(1 + |a| + |y|)s−1 ε + C(1 + (1 + |a| + |y|)/r)3(s−1) (β3 + β3s )n−1/2 . For z ∈ B(y −1 a, r + ε) \ B(y −1 a, r) we have |z| ≤ c(|a−1 y| + r + ε) consequently by (40) and (43) ∂ I |X I gyz (e)| = gyz (e) ≤ ∂η X ≤ c(I) |a−1 y|i |a−1 y|j + (r + ε)j ε−|J| (r + ε)|J|−d(J) . |J|≤|I|,i+j=d(J)−d(I)≥0 We have ε−|J| (r + ε)|J|−d(J) ≤ ε−d(I) (r + ε)d(I)−d(J) , since |J| − d(I) ≤ |J| − |I| ≤ 0. Theorem B and the above estimates imply ZZ I |X f (e)| ≤ |X I gyz (e)|δ1/√n ν k−1 (dy)δ1/√n µn−k (dz) ≤ ≤ c(I)(1 + |a|)s−1 ε + C(1 + (1 + |a|)/r)3(s−1) (β3 + β3s )n−1/2 ε−d(I) × X × (1 + |a|i ) 1 + |a|j + (r + ε)j (r + ε)d(I)−d(J) |J|≤|I|,i+j=d(J)−d(I)≥0 ≤ c(I)ε−d(I) (1 + (1 + |a|)/r)(s−1)d(I) (Sn + ε(1 + |a|)s−1 ) since |J| ≤ |I| implies d(J) ≤ s|J| ≤ s|I| ≤ sd(I). We have for I ∈ Zm + with d(I) ≤ 3 Z I |u|>1 ≤ Z |η (u)| |δ1/√n (µ − ν)|(du) ≤ 3 |u| |δ1/ √ n (µ Z − ν)|(du) ≤ cβ3 n 22 |u|>1 −3/2 . |u|d(I) |δ1/√n (µ − ν)|(du) ≤ (44) Collecting the above estimates we obtain for I ∈ Zm + with d(I) ≤ 3 Z η I (u)X I f (e)δ1/√n (µ − ν)(du) ≤ |u|>1 ≤ c(I)ε−d(I) (1 + (1 + |a|)/r)(s−1)d(I) (Sn + ε(1 + |a|)s−1 )β3 n−3/2 . Thus using ε < 1 we can conclude |Ii | ≤ cε−3 (1 + (1 + |a|)/r)3(s−1) (Sn + ε(1 + min{|a|, r})s−1 )β3 n−3/2 for i = 1, 2. By Theorem A we have |f (u) − P (2) (u)| ≤ |u|3 supd(I)=3,|x|≤b3 |u| |X I f (x)| for all u ∈ G. Applying ε < min{1, r/2} it can be proved as above that for fixed a, y, u ∈ G with |u| ≤ 1 we have δ1/√n µn−k z ∈ G : X I gyz (u) 6= 0 ≤ δ1/√n µn−k (B(u−1 y −1 a, r + ε) \ B(u−1 y −1 a, r)) 3(s−1) ≤ c(1 + |a| + |y| + |u|)s−1 ε + C (1 + (1 + |a| + |y| + |u|)/r) (β3 + β3s )n−1/2 ≤ c(1 + |a| + |y|)s−1 ε + C(1 + (1 + |a| + |y|)/r)3(s−1) (β3 + β3s )n−1/2 . Thus using (40)—(43) one can show |I3 | ≤ cε−3 (1 + (1 + |a|)/r)3(s−1) (Sn + ε(1 + |a|)s−1 )β3 n−3/2 Consequently Z ϕ(x)(γk + γ̃k )(dx) ≤ |I0 | + |I1 | + |I2 | + |I3 | ≤ ≤ cε −3 (1 + (1 + |a|)/r) 3(s−1) (Sn + ε(1 + |a|) s−1 )β3 n (45) −3/2 + cβ3 n −3/2 if ε < min{1, r/2} and 1 ≤ k ≤ [n/2]. R Now we give another estimate for ϕ(x)(γk + γ̃k )(dx) if ε < min{1, r/2} and 1 ≤ k ≤ [n/2]. First we observe Z ZZ ϕ(x)γk (dx) = l(v)pt (vu−1 )λ(dv)δ1/√n (µ − ν)(du) R where t = (k − 1)/n and l(v) = ϕ(vz)δ1/√n (µn−k − ν n−k )(dz), v ∈ G. We use the R Taylor inequality for the function f (u) = l(v)pt (vu−1 )λ(dv), u ∈ G at e ∈ G. Denote P (2) (u) the left Taylor polynomial of f at e ∈ G of homogeneous degree 2. We have again Z ϕ(x)γk (dx) = I0 − I1 − I2 + I3 23 where Z (f (u) − f (e))δ1/√n (µ − ν)(du) Z X I1 = ηi (u)Xi f (e)δ1/√n (µ − ν)(du) di =1,2 |u|>1 Z X I2 = ηi (u)ηj (u)Xi Xj f (e)δ1/√n (µ − ν)(du) di =dj =1 |u|>1 Z I3 = (f (u) − P (2) (u))δ1/√n (µ − ν)(du). I0 = |u|>1 |u|≤1 From |ϕ| ≤ 1 it follows |l| ≤ 1 and |f | ≤ 1, thus Z |I0 | ≤ 2 |δ1/√n (µ − ν)|(du) ≤ cβ3 n−3/2 . |u|>1 Put gvt (u) = pt (vu−1 ) for u, v ∈ G and t > 0. We have Z I X f (u) = l(v)X I gvt (u)λ(dv) for all x ∈ G, I ∈ Zm +. First we shall estimate I X f (e) = We have ∂ ∂η I gvt (e) = Z X l(v) ∂ ∂η I gvt (e)λ(dv). |J|≤|I|,d(J)≥d(I) PIJ (v) ∂ ∂η J pt (v) (46) where PIJ is a homogeneous polynomial of degree d(J)−d(I). Using |l| ≤ 1 and Corollary 1 we obtain Z I ∂ l(v) gvt (e)λ(dv) ≤ |v|>1 ∂η Z ∂ J X d(J)−d(I) |v| ≤ cI pt (v) λ(dv) |J|≤|I|,d(J)≥d(I) |v|>1 ∂η ≤ cI . For |v| ≤ 1 we have |l(v)| ≤ c(Sn + ε(1 + |a|)s−1 ) applying the inductive hypothesis (36) for n − k and using ε < min{1, r/2}. Corollary 1 and (46) Z I ∂ gvt (e)λ(dv) ≤ cI (Sn + ε(1 + |a|)s−1 )t−d(I)/2 . l(v) |v|≤1 ∂η 24 (47) Thus by Consequently by (44) we have for I ∈ Zm + with d(I) ≤ 3 Z η I (u)X I f (e)δ1/√n (µ − ν)(du) ≤ |u|>1 ≤ cβ3 n−3/2 + cβ3 (k − 1)−3/2 (Sn + ε(1 + |a|)s−1 ). Thus we can conclude |Ii | ≤ cβ3 (k − 1)−3/2 (Sn + ε(1 + |a|)s−1 ) + cβ3 n−3/2 for i = 1, 2. By Theorem A we have |f (u) − P (2) (u)| ≤ |u|3 supd(I)=3,|x|≤b3 |u| |X I f (x)| (48) for all u ∈ G. Using |l| ≤ 1 and Corollary 1 we have for u ∈ G with |u| ≤ 1, x ∈ G with |x| ≤ b3 |u| and for a sufficiently large constant c > 0 Z l(v)X I gvt (x)λ(dv) ≤ |v|>c Z ∂ K X d(K)−d(J) −1 ≤ cI 1 + |v| p (vx ) λ(dv) t |K|≤|J|≤|I| ∂η d(K)≥d(J)≥d(I) |v|>c Z ∂ K X d(K)−d(J) ≤ cI 1 + |w| p (w) λ(dw) t |K|≤|J|≤|I| ∂η d(K)≥d(J)≥d(I) |wx|>c ≤ cI since |x| ≤ b3 and |wx| > c imply |wx| ≤ c0 (|w| + |x|) ≤ c0 (|w| + b3 ) thus |w| ≥ −1 3 3 c−1 0 |wx| − b > c0 c − b ≥ 1 if c is sufficiently large. Moreover, for |v| ≤ c we have again (47), thus by Corollary 1 we have for u ∈ G with |u| ≤ 1 and x ∈ G with |x| ≤ b3 |u| as in the proof of Theorem 1 Z I l(v)X gvt (x)λ(dv) ≤ |v|≤c X d(K)−d(I) −d(K)/2 d(J)−d(I) −d(J)/2 . |x| t + |x| t ≤ cI (Sn + ε(1 + |a|)s−1 ) |K|≤|J|≤|I| d(K)≥d(J)≥d(I) Consequently by (44) and (48) Z |I3 | ≤ |f (u) − P (2) (u)| |δ1/√n (µ − ν)|(du) |u|≤1 ≤ cβ3 n−3/2 + c(Sn + ε(1 + |a|)s−1 ) 25 X3(s−1) j=0 β3+j (k − 1)−(3+j)/2 . Collecting the estimates we conclude Z ϕ(x)γk (dx) ≤ c(Sn + ε(1 + |a|)s−1 )(β3 + β3s )(k − 1)−3/2 + cβ3 n−3/2 . (49) We obtain as in the proof of Theorem 1 (using |l| ≤ 1) Z X3s−3 ϕ(x)γ̃k (dx) ≤ c β3+j n−(3+j)/2 ≤ c(β3 + β3s )n−3/2 j=0 for 1 ≤ k ≤ n. The same estimate is valid for R (50) ϕ(x)(γk + γ̃k )(dx) in case [n/2] ≤ k ≤ n. Let m ∈ {1, 2, . . . , [n/2]}. For k = 1, 2, . . . , m we use the estimate (45) : Xm Z ϕ(x)(γk + γ̃k )(dx) ≤ c(β3 + β3s )n−1/2 + k=1 + cε−3 (1 + (1 + |a|)/r) 3(s−1) (Sn + ε(1 + |a|)s−1 )(β3 + β3s )mn−3/2 . For k = m + 1, . . . , [n/2] we use (49) : Z ϕ(x)(γk + γ̃k )(dx) ≤ k=m+1 X[n/2] ≤ c(Sn + ε(1 + |a|)s−1 )(β3 + β3s )m−1/2 + c(β3 + β3s )n−1/2 . For k = [n/2] + 1, . . . , n we use (50) : Z ϕ(x)(γk + γ̃k )(dx) ≤ c(β3 + β3s )n−1/2 . k=[n/2]+1 Xn Let K > 0 be a large enough constant (we shall specify it later) and ε = K(1 + (1 + |a|)/r)s−1 (β3 + β3s )n−1/2 . Using (38) and collecting the above estimates we obtain δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ c (K −1 + C −1 )(K −2 (β3 + β3s )−2 m + K(β3 + β3s )m−1/2 ) + C −1 (K + 1) Sn . From Theorem 1 it follows that if β3 + β3s ≤ C1 then δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ C2 (β3 + β3s )n−1/2 ≤ C2 C −1 Sn for all n ≥ 4. 26 (51) If C1 ≤ (β3 + β3s ) < K −1 then we choose m = 1 and we have δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ c (K −1 + C −1 )(K −2 C −2 + 1) + C −1 (K + 1) Sn . 1 (52) If β3 + β3s > K −1 [n/2]1/2 then we choose m = n and we have δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ c K −1 + C −1 (K + 2) Sn . (53) If K −1 < β3 + β3s ≤ K −1 [n/2]1/2 then there exists m ∈ {1, 2, . . . , [n/2]}. such that m − 1 < K 2 (β3 + β3s )2 ≤ m and then δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ c 3K −1 + C −1 (K + 4) Sn . (54) If ε = K(1 + (1 + |a|)/r)s−1 (β3 + β3s )n−1/2 ≥ r/2 then we need a new estimate of the R integral ϕ(x)(γk + γ̃k )(dx). We use the inequality (38) with ε = 1 : Z δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ max ϕi (x)(δ1/√n µn − ν)(dx) + c(1 + |a|)s−1 , i=1,2 noting that now in the definition of Ψi one should replace ε with 1. Using again the Taylor inequality for the function ZZ f (u) = ϕ(yuz)δ1/√n ν k−1 (dy)δ1/√n µn−k (dz), u∈G at e ∈ G we obtain Z Z ϕ(x)(γk + γ̃k )(dx) ≤ |f (u) − P (2) (u)| |δ1/√n (µ − ν)|(du) where P (2) is the left Taylor polynomial of f at e ∈ G of homogeneous degree 2. We obtain Z ϕ(x)(γk + γ̃k )(dx) ≤ c(1 + (1 + |a|)/r)3(s−1) (β3 + β3s )n−3/2 . Consequently from 1 < 2r −1 ε = 2r −1 K(1 + (1 + |a|)/r)s−1 (β3 + β3s )n−1/2 we conclude δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ cC −1 (K + 1)Sn . (55) If ε = K(1 + (1 + |a|)/r)s−1 (β3 + β3s )n−1/2 ≥ 1 then we have δ1/√n (µn )(B) − ν(B) ≤ 1 ≤ K(1 + (1 + |a|)/r)s−1 (β3 + β3s )n−1/2 ≤ C −1 KSn 27 (56) From (51)—(56) it follows that δ1/√n (µn )(B(a, r)) − ν(B(a, r)) ≤ Sn if we first fix a sufficiently large K > 0, then choose a sufficiently large C > 0. The induction is complete. # COROLLARY 3. The statement of Theorem 2 remains true if we replace β 3 (µ, ν) + β3s (µ, ν) in (35) by M3 (µ, ν) + M3s (µ, ν). ACKNOWLEDGEMENT. The author is grateful to Professor Herbert Heyer and to Professor Eberhard Siebert for supporting his work. REFERENCES [1] BENTKUS, V. (1986). On the dependence on the dimension of the estimate of BerryEsseen. Lithuanian Math. J. 26 205–210. [2] BENTKUS, V. and BLOZNELIS, M. (1989). Nonuniform estimate of rate of convergence in central limit theorem with stable limit law. Lithuanian Math. J. 29 14–26. [3] BERGSTRÖM, H. (1969). On the central limit theorem in Rk . Z. Wahrsch. Verw. Gebiete 14 113–126. [4] CRÉPEL, P. and RAUGI, A. (1978). Théorème central limite sur les groupes nilpotents. Ann. Inst. H. Poincaré Probab. Statist. 14 145–164. [5] FOLLAND, G.B. and STEIN, E.M. (1982). Hardy spaces on homogeneous groups. Princeton University Press, New Jersey. [6] GUIVARC’H, Y. (1980). Sur la loi des grands nombres et le rayon spectral d’une marche aléatoire. Astérisque 74 47–98. [7] HAZOD, W. (1984). Stable and semistable probabilities on groups and on vector spaces. In: Szynal, D., Weron, A. (eds.) Probability Theory on Vector Spaces III. Proceedings, Lublin 1983. (Lect. Notes Math., vol. 1080, pp. 69–89) Springer, Berlin Heidelberg New York Tokyo. [8] HEBISCH, W. (1989). Sharp pointwise estimate for kernels of the semigroup generated by sums of even powers of vector fields on homogeneous groups. Studia Math. 95 93–106. [9] HEYER, H. (1977). Probability measures on locally compact groups. Springer, Berlin Heidelberg New York. [10] HÖRMANDER, L. (1967). Hypoelliptic second order differential equations. Acta Math. 119 147-171. [11] KUELBS, J. and KURTZ, T. (1974). Berry-Esseen estimates in Hilbert space and appli- 28 cation to the law of the iterated logarithm. Ann. Probab. 2 387–407. [12] PAP, G. (1990). Rate of convergence in CLT on Heisenberg group. (submitted to Probab. Theory Relat. Fields) [13] PAULAUSKAS, P. and RAČKAUSKAS, A. (1989). Approximation theory in the central limit theorem. Exact results in Banach spaces. Kluwer Academic Publishers, Dordrecht Boston London. [14] RAUGI, A. (1978). Théorème de la limite centrale sur les groupes nilpotents. Wahrsch. Verw. Gebiete 43 149–172. Z. [15] SAZONOV, V. (1981). Normal approximation – some recent advances. Lect. Notes Math., vol. 879, Springer, Berlin New York. [16] SIEBERT, E. (1982). Absolute continuity, singularity, and supports of Gauss semigroups on a Lie group. Monatsh. Math. 93 239-253. MATHEMATICAL INSTITUTE LAJOS KOSSUTH UNIVERSITY PF. 12 H-4010 DEBRECEN, HUNGARY 29