International Mathematical Forum, 3, 2008, no. 15, 721 - 738 Berry-Esseen Bound for Independent Random Sum via Stein’s Method N. Chaidee and K. Neammanee Department of Mathematics, Faculty of Science Chulalongkorn University, Bangkok 10330, Thailand [email protected] Abstract Let (Xn) be a sequence of independent identically distributed random variables with zero means and finite third moments and N a positive integral-valued random variable such that N, X1, X2 , . . . are independent. In this work, we apply a new technique, which is called Stein’s method, to give uniform and non-uniform bounds in normal approximation of a random sum X1 + X2 + · · · + XN . Mathematics Subject Classification: 60F05, 60G05 Keywords: normal approximation, Stein’s method, random sum, concentration inequality, uniform and non-uniform bounds 1 Introduction One of the most important theorem in probability theory and statistics is the central limit theorem(CLT) which is stated as follows. Let X1 , X2 , . . . , Xn be independent random variables with EXi = 0 and EXi2 = σi2 < ∞. Then for Wn = X1 + X2 + · · · + Xn s2n and s2n = σ12 + σ22 + · · · + σn2 we have P (Wn ≤ x) → Φ(x) as n → ∞ where Φ(x) is the standard normal distribution function, that is, x 1 2 Φ(x) = √ e−t /2 dt. 2π −∞ 722 N. Chaidee and K. Neammanee To apply the CLT we always need to know a bound between P (Wn ≤ x) and Φ(x), i.e., |P (Wn ≤ x) − Φ(x)|. Berry([2]) and Esseen([6]) are the first two mathematicians who gave this bound. Their bound is of the form n C E|Xi |3 . sup |P (Wn ≤ x) − Φ(x)| ≤ 3 s x∈Ê n i=1 (1) Many authors([2], [6], [20]) find and improve the constant C in (1) and the best C is 0.7655 which is obtained by Shiganov in 1986([12]) in case of X1 , X2 , . . . , Xn are identically and 0.7915 in case of non-identically([12]). In many situations, we need to approximation the probability of random sum. Let X1 , X2 , . . . be independent identically distributed random variables with EXi = 0 and E|Xi |3 = γ < ∞. Let N be a positive integral-valued random variable with EN < ∞ and N, X1 , X2 , . . . are independent. Let W = X1 + X2 + · · · + XN √ EN σ where σ 2 = EXi2 for all i = 1, 2, . . . . The CLT for W are obtained by many mathematicians(see, [14], [1], [22], [9], [13], [15], [11] and [5] for examples). In this work, we concerns on uniform and non-uniform bounds between P (W ≤ x) and Φ(x) by using a new method, which is called Stein method. Normally, the method which is used in our literatures is Fourier transform while in this paper we will use a new method which introduced by Stein in 1972([7]) to obtain bounds of |P (W ≤ x) − Φ(x)|. Stein’s method was introduced by Stein in 1972 for normal approximation of dependent random variables. This method use differential equation which always call Stein’s equation in stead of Fourier transform. There are three approaches of Stein’s method for normal approximation, namely, inductive approach(for example, see [4], [10]), coupling approach(for example, see [16], [8]) and concentration inequality approach(for example, see [21]). The concentration inequality approach was originally used by Stein for independent and identically distributed random variable([21]) and it was extended by Chen in([18], [19], [17]). In this paper, we use concentration inequality approach which used in [18] to find bounds of |P (W ≤ x) − Φ(x)|. A uniform bound in normal approximation is in section 2 while a non-uniform bound is in section 3. 2 Uniform Bound Let X1 , X2 , . . . be independent identically distributed random variables with EXi = 0, EXi2 = σ 2 and E|Xi |3 = γ < ∞. Let N be an integral-valued 723 Berry-Esseen bound for independent random sum random variable with EN 3/2 < ∞ and N, X1 , X2 , . . . are independent. Define Xi Yi = √ EN σ and W = Y1 + Y2 + · · · + YN . Then EYi = 0 and 2 EW = E( N i=1 2 Yi ) = ∞ P (N = n) n=1 n EYi2 = 1. i=1 In this section, we provide a uniform bound between P (W ≤ x) and Φ(x). Our main result is in Theorem 2.1. Theorem 2.1. There exists a constant C such that 1.5δEN 3/2 E|N − EN| 0.125 + + sup |P (W ≤ x) − Φ(x)| ≤ 10.32δ + √ (EN)3/2 EN EN x∈Ê γ . where δ = √ EN σ 3 Corollary 2.2. If N = n, then there exists a constants C such that sup |P (W ≤ x) − Φ(x)| ≤ x∈ Ê 1 11.82γ √ √ (1 + ). σ3 n n To prove Theorem 2.1, we need the following concentration inequality. Proposition 2.3. (Uniform Concentration Inequality). Let Wn = n Yi and Wn(i) = Wn − Yi for i = 1, 2, . . . , n. i=1 If δ < 1 , then for a < b 20 P (a ≤ Wn(i) 2.5EN 1 { (b − a) + δ} ≤ b) ≤ n 2 δ n + 2δ + √ . EN EN 724 N. Chaidee and K. Neammanee Proof. In [18], Chen and Shao defined ⎧ 1 ⎪ ⎨− 2 (b − a) − δ f(t) = t − 12 (b + a) ⎪ ⎩1 (b − a) + δ 2 f : R → R by for t < a − δ, for a − δ ≤ t ≤ b + δ, for t ≥ b + δ and showed that EWn(i) f(Wn(i) ) ≥ E{I(a ≤ Wn(i) ≤ b)S} − P (a ≤ Wn(i) ≤ b)E(|Yi | min{δ, |Yi|}), (2) n n EXj2 1 E|Xj |3 − ES ≥ ENσ 2 2δ j=1 (ENσ 2 )3/2 j=1 (3) 1 |EWn(i) f(Wn(i) )| ≤ { (b − a) + δ}E|Wn(i) | 2 (4) and where S= n |Yj | min{δ, |Yj |}, j=1 I(A)(w) = 1 if w ∈ A, 0 if w ∈ / A. By (3), ES ≥ nγ n n − . = 2 3/2 EN 2δ(ENσ ) 2EN (5) From this fact and the fact that n 2 ES = E( |Yj | min{δ, |Yj |})2 = j=1 n E(|Yj | min{δ, |Yj |})2 + j=1 ≤ n E(|Yj | min{δ, |Yj |})2 + ( n j=1 2 = E[|Yi | min{δ, |Yi|}][|Yj | min{δ, |Yj |}] i,j i=j n j=1 ≤ δ2 j=1 E|Yj |2 + E 2 S nδ + E 2 S, EN E|Yj | min{δ, |Yj |})2 Berry-Esseen bound for independent random sum 725 we have P (S ≤ 0.4n 0.4n ) = P (E(S) − S ≥ E(S) − ) EN EN n ≤ P (E(S) − S ≥ ) 10EN 100(EN)2 E(S − E(S))2 ≤ n2 100(EN)2 = (ES 2 − E 2S) n2 5δEN . ≤ n Hence, by (2) and the fact that for s < t, y ≥ 0,c > 0, I(s ≤ w ≤ t)y ≥ c(I(s ≤ w ≤ t) − (1 − y/c)I(y ≤ c)), (6) we have 0.4n 2.5SEN 0.4n {P (a ≤ Wn(i) ≤ b) − E[(1 − )I(S ≤ )]} EN n EN − P (a ≤ Wn(i) ≤ b)E(|Yi | min{δ, |Yi |}) 0.4n 0.4n ≥ {P (a ≤ Wn(i) ≤ b) − EI(S ≤ )} − δE|Yi| EN EN δ 0.4n 0.4n {P (a ≤ Wn(i) ≤ b) − P (S ≤ )} − √ ≥ EN EN EN 0.4n δ P (a ≤ Wn(i) ≤ b) − 2δ − √ ≥ . EN EN EWn(i) f(Wn(i) ) ≥ This implies, 2.5EN δ EWn(i) f(Wn(i) ) + 2δ + √ n EN δ 2.5EN 1 (i) ( (b − a) + δ) E|Wn |2 + 2δ + √ ≤ n 2 EN 2.5EN 1 δ n ( (b − a) + δ) ≤ + 2δ + √ n 2 EN EN P (a ≤ Wn(i) ≤ b) ≤ where we have used (4) in the second inequality. Proof of Theorem 2.1. To bound P (W ≤ x) − Φ(x), it suffices to consider x ≥ 0 as we can apply result to −W when x < 0. By [18], p. 246, we have |P (W ≤ x) − Φ(x)| ≤ 0.55 for x ≥ 0, 726 N. Chaidee and K. Neammanee hence we only need to consider the case where δ< 1 . 20 Let f be the unique bounded solution fx of the Stein equation f (w) − wf (w) = I(w ≤ x) − Φ(x). (7) If we replace w in (7) by Wn , then Ef (Wn ) − EWn f(Wn ) = P (Wn ≤ x) − Φ(x). Set pn = P (N = n). Then, by the fact that ∞ n EWn f(Wn ) = E f (Wn(i) + t)Mi(t)dt −∞ i=1 (eq. (2.17), p.12 of [3]) where Mi (t) = E[Yi (I(0 ≤ t ≤ Yi ) − I(Yi ≤ t < 0))], yields P (W ≤ x) − Φ(x) ∞ ∞ pn P (Wn ≤ x) − pn Φ(x) = n=1 = = ∞ n=1 ∞ n=1 pn [P (Wn ≤ x) − Φ(x)] pn [Ef (Wn ) − EWn f(Wn )] n=1 = ∞ pn n=1 ∞ + n ∞ E −∞ i=1 f (Wn )Mi (t)dt − pn Ef (Wn ) − n n=1 −∞ ∞ E −∞ i=1 ∞ E i=1 n f (Wn(i) + t)Mi (t)dt f (Wn )Mi (t)dt = A1 + A2 (8) where A1 = ∞ pn n=1 n i=1 ∞ E −∞ f (Wn )Mi (t)dt − ∞ n pn Ef (Wn ) − E A2 = n=1 i=1 n i=1 ∞ −∞ ∞ E −∞ f (Wn(i) + t)Mi (t)dt , f (Wn )Mi (t)dt . 727 Berry-Esseen bound for independent random sum By noting that ⎧ ⎪ if w + s ≤ x, w + t > x, ⎨1 f (w + s) − f (w + t) ≤ (|w| + 0.63)(|s| + |t|) if s ≥ t, ⎪ ⎩ 0 otherwise (see [18], p. 247), we have A1 = ≤ ∞ n=1 ∞ n pn E n n=1 ≤ ≤ ∞ pn n n=1 i=1 + pn ∞ E (i) Wn +t>x (i) Wn +Yi ≤x E E Yi ≥t n=1 i=1 n ∞ EE Yi + ∞ n=1 ∞ pn pn n i=1 n n=1 i=1 n=1 i=1 ∞ E i=1 E Yi ≥t (|Wn(i) | + 0.63)(|Y i| + |t|)Mi(t)dt (|Wn(i) | + 0.63)(|Y i| + |t|)Mi(t)dt t>Yi i=1 Mi (t)dt + n I(x − t < Wn(i) ≤ x − Yi )Mi (t)dt t>Yi n pn f (Wn(i) + Yi ) − f (Wn(i) + t) Mi (t)dt i=1 n=1 ≤ −∞ i=1 pn ∞ I(x − t < Wn(i) ≤ x − Yi )Mi (t)dt E(|Wn(i) | + ∞ 0.63)E −∞ (|Yi | + |t|)Mi (t)dt E Yi I(x − t < Wn(i) ≤ x − Yi )Mi (t)dt t>Yi n 1 (i) pn ( E(Wn )2 + 0.63)(E |Yi|E|Yi|2 + E|Yi |3 ) + 2 n=1 i=1 ∞ n pn E P (x − t < Wn(i) ≤ x − Yi |Yi )Mi (t)dt ≤ ∞ 3 + 2 ≤ ∞ n=1 t>Yi pn n=1 n pn i=1 n ( i=1 E ∞ −∞ n + 0.63)E |Yi|3 EN P (x − t < Wn(i) ≤ x − Yi |Yi )Mi (t)dt ∞ 3γ pn n( + 2(EN )3/2σ 3 n=1 n + 0.63) EN 728 N. Chaidee and K. Neammanee 2.5EN 1 δ n ≤ Mi (t)dt { (|t| + |Yi |) + δ} + 2δ + √ pn E n 2 EN EN t>Yi n=1 i=1 ∞ n 0.945γEN 1.5γEN 3/2 + 2 3 (EN ) σ (EN )3/2σ 3 n ∞ 2.5EN 1 E|Yi|3 n 2 2 ≤ pn { ( + E|Yi|E|Yi| ) + δE|Yi| } + 2δE|Yi|2 n 2 2 EN n=1 i=1 δE|Yi|2 1.5δEN 3/2 + √ + 0.945δ + (EN )3/2 EN + 2.5δ 1.5δEN 3/2 = 9.375δ + √ + + 0.945δ (EN )3/2 EN 1.5δEN 3/2 0.125 + ≤ 10.32δ + √ (EN )3/2 EN (9) where we have used Proposition 2.3 in the seventh Similar to (9), we can use the fact that ⎧ ⎪ ⎨−1 f (w + s) − f (w + t) ≥ −(|w| + 0.63)(|s| + |t|) ⎪ ⎩ 0 inequality. if w + s > x, w + t ≤ x, if s < t, otherwise, (see [18], p. 247) to prove that 1.5δEN 3/2 0.125 A1 ≥ −10.32δ − √ − . (EN)3/2 EN Hence 1.5δEN 3/2 0.125 |A1| ≤ 10.32δ + √ + . (EN)3/2 EN By [18], p. 247, |f (w)| ≤ 1 for all w, (10) then ∞ n |A2| = pn Ef (Wn ) 1 − E ≤ n=1 ∞ n n=1 i=1 pn 1 − EYi2 n = pn 1 − EN n=1 ∞ i=1 ∞ −∞ Mi (t)dt Berry-Esseen bound for independent random sum N = E 1 − EN E|N − EN| = . EN By combining (8)-(11), we have 1.5δEN 3/2 E|N − EN| 0.125 √ + + |P (W ≤ x) − Φ(x)| ≤ 10.32δ + . (EN)3/2 EN EN 3 729 (11) 2 Non-uniform Bound In this section, we will give a non-uniform bound in normal approximation of W = Y1 + Y2 + · · · + YN and further assume that EN 5 < ∞. Theorem 3.1. There exists a constant C such that for every real number x, EN 5 C (V arN )EN 2 Cδ 1+ + . |P (W ≤ x) − Φ(x)| ≤ (1 + |x|)3 (EN)5 (1 + |x|)2(EN)2 Corollary 3.2. If N = n, then there exists a constants C such that for every real number x, Cδ |P (W ≤ x) − Φ(x)| ≤ . (1 + |x|)3 The following proposition concerns a concentration inequality for a nonuniform bound. Proposition 3.3. (Non-uniform Concentration Inequality) For 0 ≤ a < b, we have n4 C(b − a + δ) EN + . P (a ≤ Wn(i) ≤ b) ≤ (1 + a)3 n (EN)4 where C is a constant. Proof. In the proof of non-uniform concentration inequality in [18], p. 241, Chen and Shao defined Zj = Yj I(|Yj | ≤ 1 + a) − EYj I(|Yj | ≤ 1 + a), n Zj and Tn(i) = r= j=1 j=i n j=1 j=i EYj I(|Yj | > 1 + a), 730 N. Chaidee and K. Neammanee and showed that P (a ≤ Wn(i) ≤ b) ≤ P (a + r ≤ Tn(i) ≤ b + r) + P ( max |Yj | > 1 + a). 1≤j≤n Hence, by the fact that n 1 nδ P ( max |Yj | > 1 + a) ≤ E|Yj |3 = , 3 1≤j≤n (1 + a) j=1 EN(1 + a)3 we have P (a ≤ Wn(i) ≤ b) ≤ P (a + r ≤ Tn(i) ≤ b + r) + nδ . EN(1 + a)3 (12) To prove this proposition, it suffices to bound P (a + r ≤ Tn(i) ≤ b + r). By Rosenthal inequality(see, for example, [23], p.59) and the fact that n E|Tn(i) |2 ≤ , EN we have n (i) 4 (i) 2 2 E|Yj |4I(|Yj | ≤ 1 + a) E|Tn | ≤ C (E|Tn | ) + (13) j=1 j=i 2 2 3 ≤ C (nEY1 ) + n(1 + a)E|Y1 | n2 n(1 + a)δ + . ≤C (EN)2 EN (14) Case 1. (1 + a)δ ≥ 1. By the fact that r ≤ n 1 and < δ, we have EN 1+a P (a + r ≤ Tn(i) ≤ b + r) ≤ P (1 + a ≤ 1 − r + Tn(i) ) (i) ≤ E|1 − r + Tn |4 (1 + a)4 (i) ≤ ≤ ≤ ≤ ≤ C(1 + r 4 + E|Tn |4 ) (1 + a)4 n2 C Cn4 C n(1 + a)δ + + + (1 + a)4 (EN )4(1 + a)4 (1 + a)4 (EN )2 EN 2 4 n n n Cδ + + 1+ 3 2 (1 + a) EN (EN ) (EN )4 n4 Cδ {1 + } 3 (1 + a) (EN )4 n4 EN Cδ + { }. (15) (1 + a)3 n (EN )4 731 Berry-Esseen bound for independent random sum Case 2. (1 + a)δ < 1. n−1 Follows the argument in [18], p. 243 and choose C in (6) is , we can 4EN show that E(Tn(i) f(Tn(i) )) ≥ where U = n (n − 1)(1 + a)3 n−1 P (a + r ≤ Tn(i) ≤ b + r) − P (U ≤ ) 4EN 4EN (16) ηj and ηj = |Zj | min(δ, |Zj |) and j=1 j=i ⎧ ⎪ for x < a + r − δ, ⎨0 f(x) = (1 + x − r + δ)3(x − a − r + δ) for a + r − δ ≤ x ≤ b + r + δ, ⎪ ⎩ for x > b + r + δ. (1 + x − r + δ)3(b − a + 2δ) d2 and EYi = 0, we have By the fact that for any c, d > 0, min(c, d) ≥ d − 4c EU ≥ = n |Zj |3 E Zj2 − 4δ j=1 j=i n EYj2I(|Yj | ≤ 1 + a) − (EYj I(|Yj | ≤ 1 + a))2 j=1 j=i 732 N. Chaidee and K. Neammanee − = n 3 1 E Yj I(|Yj | ≤ 1 + a) − EYj I(|Yj | ≤ 1 + a) 4δ j=1 j=i n EYj2I(|Yj | ≤ 1 + a) − (EYj I(|Yj | > 1 + a))2 j=1 j=i n 3 1 E Yj I(|Yj | ≤ 1 + a) − EYj I(|Yj | ≤ 1 + a) − 4δ j=1 ≥ n j=i EYj2 I(|Yj | ≤ 1 + a) − EYj2 I(|Yj | > 1 + a) j=1 j=i 1 3 3 E |Yj | I(|Yj | ≤ 1 + a) + |EYj I(|Yj | ≤ 1 + a)| − δ j=1 n j=i ≥ n j=1 j=i EYj2 2 − E|Yj |3 δ j=1 n j=i 2 (n − 1)σ 2(n − 1)γ − ENσ 2 δ(ENσ 2)3/2 n−1 = . 2EN = From this fact and Rosenthal inequality, P (U ≤ n−1 n−1 n−1 ) ≤ P (EU − U ≥ − ) 4EN 2EN 4EN n−1 ) = P (EU − U ≥ 4EN 16(EN)2 ≤ E|U − E(U)|4 (n − 1)2 n n C(EN)2 2 2 4 ≤ Eηj + Eηj (n − 1)2 j=1 j=1 2 j=i j=i 2 C(EN) 3 4 4 (n − 1)δE|Z1 | + (n − 1)δ E|Z1 | ≤ (n − 1)2 2 C(EN)2 3 4 3 + (n − 1)(1 + a)δ E|Y1 | (n − 1)δE|Y1 | ≤ (n − 1)2 Berry-Esseen bound for independent random sum C(EN)2 (n − 1)δ 2 2 (n − 1)(1 + a)δ 5 + (n − 1)2 EN EN 5 (1 + a)δ EN = C δ4 + n−1 Cδ EN ≤ 1 + (1 + a)3 n−1 Cδ EN ≤ 1 + (1 + a)3 n 733 = (17) where we have used the fact that (1 + a)δ < 1 in the last inequality. By (13), (14), (16) and (17), P (a + r ≤ Tn(i) ≤ b + r) 4EN n−1 )+ ≤ P (U ≤ E(Tn(i) f(Tn(i) )) 3 4EN (n − 1)(1 + a) EN Cδ CEN {(b − a + 2δ)E|Tn(i) (1 + Tn(i) − r + δ)3|} 1 + ≤ + 3 3 (1 + a) n n(1 + a) CEN EN Cδ + 1 + ≤ {(b − a + 2δ)((1 + r3 )E|Tn(i) | + E|Tn(i) |4)} (1 + a)3 n n(1 + a)3 EN Cδ 1 + ≤ (1 + a)3 n CEN n3 n2 n(1 + a)δ (i) 2 + (b − a + δ){1 + E|T | + + } n n(1 + a)3 (EN)3 (EN)2 EN EN Cδ 1 + ≤ (1 + a)3 n √ CEN n n7/2 n2 n + + (b − a + 2δ){ √ + } + n(1 + a)3 EN (EN)2 (EN)7/2 EN EN CEN(b − a + δ) n4 Cδ + 1 + 1 + ≤ (1 + a)3 n n(1 + a)3 (EN)4 n3 C(b − a + δ) EN + . (18) ≤ (1 + a)3 n (EN)3 By (12), (15) and (18), we have P (a ≤ Wn(i) C(b − a + δ) EN n4 ≤ b) ≤ . + (1 + a)3 n (EN)4 734 N. Chaidee and K. Neammanee Lemma 3.4. Let fx be the solution of (7) and g(w) = (wfx (w)) . 1. For x > 0, there exists a constant C such that E|fx (Wn )| ≤ n C . 1 + (1 + x)2 EN x 2. For x ≥ 4 and |u| ≤ 1 + , we have 4 EN C n4 (1 + δx) + Eg(Wn(i) + u) ≤ (1 + x)3 n (EN)4 for i = 1, 2, . . . , n. Proof. 1. Using the argument of Lemma 5.1 in [18], p.248 and the fact that EWn2 = n , EN we have E|fx (Wn )| ≤ C n . 1 + (1 + x)2 EN 2. Use the argument of Lemma 5.2 in [18], p.248 and Proposition 3.3. Proof of Theorem 3.1. In view of Theorem 2.1, we may without loss of generality, assume x ≥ 4. If (1 + x)δ ≥ 1, then, the arguments in (14) and (15) yield P (W ≥ x) ≤ EN 4 Cδ 1 + (1 + x)3 (EN)4 which implies |F (x) − Φ(x)| ≤ P (W ≥ x) + 1 − Φ(x) EN 4 C Cδ 1 + + ≤ 3 4 (1 + x) (EN) (1 + x)4 EN 4 Cδ 1 + . ≤ (1 + x)3 (EN)4 Hence, we assume (1 + x)δ < 1. Berry-Esseen bound for independent random sum 735 By the same argument as (8), we have F (x) − Φ(x) ∞ n ∞ x pn E I(|Yi | ≤ 1 + ) (f (Wn(i) + Yi ) − f (Wn(i) + t))Mi (t)dt = 4 −∞ n=1 i=1 ∞ n ∞ x pn E I(|Yi| > 1 + ) (f (Wn(i) + Yi ) − f (Wn(i) + t))Mi(t)dt + 4 −∞ n=1 i=1 ∞ ∞ n pn Ef (Wn ) − E f (Wn )Mi (t)dt + n=1 i=1 −∞ := A1 + A2 + A3. One can follows the argument of [18] on p.251-252 to show that Yi ∞ n ∞ x pn Mi (t) Eg(Wn(i) + u)dudt |A1| ≤ E I(|Yi| ≤ 1 + ) 4 −∞ t n=1 i=1 n ∞ x pn E I(|Yi| ≤ 1 + ) + 4 n=1 i=1 ∞ × P (x − max(Yi , t) < Wn(t) ≤ x − min(Yi , t)|Yi)Mi (t)dt −∞ = A11 + A12 . Then Proposition 3.3 and Lemma 3.4(2) yields |A11 | ≤ ≤ ≤ = = = n ∞ C x pn E I(|Yi | ≤ 1 + ) 3 (1 + x) n=1 i=1 4 ∞ Yi EN n4 × dudt Mi (t) (1 + δx) + 4 n (EN) −∞ t n ∞ n4 C(1 + δx) EN + pn E|Yi |3 (1 + x)3 n=1 n (EN)4 i=1 ∞ EN n4 nδ C + p n (1 + x)3 n=1 n (EN)4 EN ∞ n5 Cδ EN + p n (1 + x)3EN n=1 (EN)4 EN 5 Cδ EN + (1 + x)3EN (EN)4 EN 5 Cδ 1+ (1 + x)3 (EN)5 736 N. Chaidee and K. Neammanee where we have used the fact that (1 + x)δ < 1 in the third inequality, and ∞ EN n4 C + |A12| ≤ pn (1 + x)3 n=1 n (EN)4 n x × E I(|Yi | ≤ 1 + ) (|t| + |Yi | + δ)Ki (t)dt 4 |t|≤1+ x4 i=1 n ∞ EN C n4 pn (E|Yi |3 + δE|Yi |2) ≤ + (1 + x)3 n=1 n (EN)4 i=1 ∞ EN C n4 nδ nδ ( pn = + + ) 3 4 (1 + x) n=1 n (EN) EN EN ∞ n5 Cδ pn EN + = (1 + x)3 EN n=1 (EN)4 Cδ EN 5 = 1+ . (1 + x)3 (EN)5 Hence |A1| ≤ EN 5 Cδ 1 + . (1 + x)3 (EN)5 By (10), we have |A2| ≤ ∞ pn n=1 ∞ ≤C n=1 n x P (|Yi | > 1 + ) 4 i=1 pn n E|Yi |3 (1 + x)3 i=1 ∞ C nδ ≤ pn 3 (1 + x) n=1 EN ≤ Cδ . (1 + x)3 Follows the argument of (11) and Lemma 3.4(1), we have n n C n |A3| ≤ 1 + p − 1 n (1 + x)2 i=1 EN EN CE|N 2 − (EN)2 | (1 + x)2 (EN)2 C (V arN )EN 2 ≤ . (1 + x)2 (EN)2 = Berry-Esseen bound for independent random sum 737 Hence, EN 5 C (V arN )EN 2 Cδ 1+ + . |P (W ≤ x) − Φ(x)| ≤ (1 + x)3 (EN)5 (1 + x)2(EN)2 2 References [1] A. Krajka and Z. Rychilk, Necessary and sufficient conditions for weak convergence of random sums of independent random variables. Comment.Math.Univ.Carolin., 34(1993), 465-482. [2] A.C. Berry, The accuracy of the Gaussian approximation to the sum of independent variables. Trans.Amer.Math.Soc., 49(1941), 122-136. [3] A.D. Barbour and L.H.Y.Chen, An introduction to Stein’s method, Lecture Notes Series, Institute for Mathematical Sciences, National University of Singapore, 4(2005), Singapore university press and world scientific publishing, Singapore. [4] A.D. Barbour and P. Hall, Stein’s method and the Berry-Esseen theorem. Austral. J. Statist., 26(1984), 8-15. [5] A.V. Pecinkin, The convergence to normal law of sums of random number of random summands. Theor.Verojatnost. i Primenen., 18(1973), 380382. [6] C.G. Esseen, Fourier analysis of distribution functions: a mathematical study of the Laplace-Gaussian law, Acta Math., 77(1945), 1-125. [7] C. Stein, A bounded on the error in the normal approximation to a distribution of sums of dependent random variables. Proc.Sixth Berkeley Symp.Math.Statist.Prob., 7(1972), 583-602. [8] C. Stein, Approximate computation of expectations. Institute of Mathematical Statistics, Lecture notes-monograph series 7, Hayward, California, 1986. [9] D. Landers and L. Rogge, The exact approximation order in the central limit theorem for random summation. Z.Wahrsch.Verw.Gebiete., 36(1976), 269-283. [10] E. Bolthausen, An estimate of the remainder in a combinatorial central limit theorem. z.Wahrsch.Verw.Gebiete., 66(1984), 379-386. 738 N. Chaidee and K. Neammanee [11] G. Orazov, Refinement of theorems on the saymptotic distribution of sums of a random number of random summands under various normalizations. Taskent.Gos.Univ.Nauch.Trudy Vyp., 402(1972), 87-93. [12] I.S. Shiganov, A refinement of the upper bound of the constant in the remainder term of the central limit theorem. J. Soviet Math.,3(1986), 2545-2550. [13] J. Mogyorodi, A central limit theorem for the sums of a random number of random variables. Ann.Univ.Sci.Budapest.Eotvos Sect.Math., 10(1967), 171-182. [14] J.R. Blum, D.L. Handon and J.I. Rosenblatt, On the central limit theorem for the sum of a random number of independent random variables. Z.Wahrsch.Verw.Gebiete.,1(1962), 389-393. [15] K. Neammanee, A limit theorem for random sums of independent random variables with finite variance. Doctoral dissertation, Department of Mathematics, Graduate School, Chulalongkorn University, 1993. [16] K. Neammanee, On the constant in the non-uniform version of the BerryEsseen theorem. Int.J.Math. and Math.Sci.,12(2005), 1951-1967. [17] L.H.Y. Chen, Stein’s method: some perspectives with applications, Probability Towards 2000(L.Accardi and C.C. Heyde, eds.), Lecture Notes in Statistics, 128(1998), Springer Verlag, 97-122. [18] L.H.Y. Chen and Q.M. Shao, A non-uniform Berry-Esseen bound via Stein’s method. Probab. Theory Related Fields, 120(2001), 236-254. [19] L.H.Y. Chen and Q.M. Shao, Normal approximation under local dependence. Ann.Probab., 32(2004), 1985-2028. [20] P. Van Beeck, An application of Fourier methods to the problem of sharpening the Berry-Esseen inequality. Z.Wahrsch.Verw.Gebiete., 23(1972), 187-196. [21] S.T. Ho and L.H.Y. Chen, An Lp bound for the remainder in a combinatorial central limit theorem. Ann. Probab., 6(1978), 231-249. [22] V.M. Kruglov, The convergence of the distributions of sums of a random number of independent random variables to the normal distribution. Vestnik Moskov.Univ.Ser.I Mat.Meh., 31(1976), 5-12. [23] V.V. Petrov, Limit theorems of probability theory : sequences of independent random variables. Oxford Studies in Probability 4, Clarendon Press, Oxford, 1995.
© Copyright 2024 Paperzz