• No results found

A Limit Theorem for the Moment of Self Normalized Sums

N/A
N/A
Protected

Academic year: 2020

Share "A Limit Theorem for the Moment of Self Normalized Sums"

Copied!
10
0
0

Loading.... (view fulltext now)

Full text

(1)

Journal of Inequalities and Applications Volume 2009, Article ID 957056,10pages doi:10.1155/2009/957056

Research Article

A Limit Theorem for the Moment of

Self-Normalized Sums

Qing-pei Zang

Department of Mathematics, Huaiyin Teachers College, Huaian 223300, China

Correspondence should be addressed to Qing-pei Zang,[email protected]

Received 25 December 2008; Revised 30 March 2009; Accepted 18 June 2009

Recommended by Jewgeni Dshalalow

Let{X, Xn;n ≥ 1} be a sequence of independent and identically distributed i.i.d. random variables andXis in the domain of attraction of the normal law andEX0. For 1≤p <2, b >−1, we prove the precise asymptotics in Davis law of large numbers for∞n1lognb/nE{|Sn|/Vn

ε2 logn2−p/2p}asε0.

Copyrightq2009 Qing-pei Zang. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. Introduction and Main Result

Throughout this paper, we let{X, Xn;n≥1}be a sequence ofi.i.d.random variables andXis

in the domain of attraction of the normal law andEX0. Put

Sn n

k1

Xk, Vn2 n

i1

X2i. 1.1

Also let lognlnne.Then by the well-known Davis laws of large numbers1,

n1 logn

n P

|Sn| ≥ε

nlogn

<, ε >0, 1.2

(2)

Gut and Sp˘ataru2proved its precise asymptotics as follows.

Theorem A. Suppose thatEX10andEX21 σ2<.Then for0≤δ≤1,

lim

ε0ε

2δ1∞

n1

log

n P

|Sn| ≥ε

nlogn

μ2δ2

δ1 σ

2δ2, 1.3

whereμ2δ2stands for the2δ2thabsolute moment of the standard normal distribution.

It is well known that, fori.i.d. random variables, Chow3discussed the complete moment convergence, and got the following result.

Theorem B. Let{X, Xn;n ≥ 1}be a sequence ofi.i.d.random variables with EX1 0 . Assume

p≥1,α >1/2, pα >1, andE|X|p|X|log1|X|<.Then for anyε >0,

n1

npα−2−αE max

jn |Sj| −εn α

<. 1.4

On the other hand, the past decade has witnessed a significant development on the limit theorems for the so-called self-normalized sumSn/Vn, Vn

n

i1Xi2. Bentkus and

G ¨otze 4obtained Berry-Esseen inequalities for self-normalized sums. Wang and Jing 5

derived exponential nonuniform Berry-Esseen bound. Gin´e et al.6, established asymptotic normality of self-normalized sums.

Theorem C. Let{X, Xn;n≥1}be a sequence ofi.i.d.random variables withEX1 0. Then for any

x∈R,

lim

n→ ∞P

Sn

Vnx

Φ x 1.5

holds, if and only ifXis in the domain of attraction of the normal law, whereΦxis the distribution function of the standard normal random variable.

Shao 7 showed a self-normalization large deviation result for PSn/Vnxn

without any moment conditions.

Theorem D. Let{xn;n≥1}be a sequence of positive numbers withxn → ∞andxn onas

n → ∞. IfEX0andEX2I|X| ≤xis slowly varying asx → ∞,then

lim

n→ ∞x −2

n lnP

Sn

Vnxn

−1

2. 1.6

(3)

Inspired by the above results, in this note we study the precise asymptotics in Davis law of large numbers for the moment of self-normalized sums. Our main result is as follows.

Theorem 1.1. SupposeX is in the domain of attraction of the normal law and EX 0. Then, for b >−1and1≤p <2, one has

lim

ε0ε

2pb1/2−p

n1

lognb

n E

|Sn|

Vnε

2 logn2−p/2p

2−b−1

2−p

b12pbp2E|N|

2pbp2/2−p,

1.7

here and in the sequel,Nis the standard normal random variable.

Remark 1.2. Ifp 1 and 0 < σ2 EX2 < ∞, by the strong law of large numbers, we have V2

n/nσ2, a.s.Then, we can easily obtain the following result:

lim

ε0ε

2b1∞

n1

lognb

n3/2 E |Sn| −εσ

2nlogn

σ2−b−1

b1 2b3E|N|

2b3. 1.8

Remark 1.3. As is well known, the strong approximation method is taken in order to obtain such an analogous result, however, this method is not applicable here.

2. Proof of

Theorem 1.1

In this section, we set expM/ε2p/2−p,forM >1 andε >0. Here and in the sequel,

Cwill denote positive constants, possibly varying from place to place, andx means the largest integer≤x. The proof ofTheorem 1.1is based on the following propositions.

Proposition 2.1. Forb >−1, one has

lim

ε0ε

2pb1/2−p

n1

lognb

n E

|N| −ε2 logn2−p/2p

2−b−1

2−p

b12pbp2E|N|

2pbp2/2−p.

(4)

Proof. Via the change of variable2 logt2−p/2p, we have

lim

ε0ε

2pb1/2−p

n1

lognb

n E

|N| −ε2 logn2−p/2p

lim

ε0ε

2pb1/2−p

n1

lognb n

ε2 logn2−p/2p

P|N| ≥xdx

lim

ε0ε

2pb1/2−p

e

logtb

t

ε2 logt2−p/2p

P|N| ≥xdxdt

lim

ε0

p2−b

2−p

ε22−p/2py

2p/2−pb1−1

y

P|N| ≥xdxdy

lim

ε0

p2−b

2−p

ε22−p/2pP|N| ≥x

x

ε22−p/2py

2p/2−pb1−1dydx

lim

ε0 2−b−1

b1

ε22−p/2pP|N| ≥x

x2p/2−pb1−ε2p/2−pb1·2b1dx

lim

ε0 2−b−1

b1

ε22−p/2px

2p/2−pb1P|N| ≥xdx

2−b−1

2−p

b12pbp2E|N|

2pbp2/2−p.

2.2

Proposition 2.2. Forb >−1, one has

lim

ε0ε

2pb1/2−p n

lognb

n

E |Sn|

Vnε

2 logn2−p/2p

E

|N|−ε2 logn2−p/2p

0.

2.3

Proof. SetΔnsupxR|P|Sn|/VnxP|N| ≥x|.Then, by1.5, it is easy to seeΔn → 0

asn → ∞.Observe that

lim

ε0ε

2pb1/2−p

n

lognb

n

E |Sn|

Vnε

2 logn2−p/2p

E

|N| −ε2 logn2−p/2p

lim

ε0ε

2pb1/2−p n

lognb

n × 0 P |

Sn|

Vn

2 logn2−p/2p

dx

0

(5)

≤lim

ε0ε

2pb1/2−p

n

lognb

n

0 P

|

Sn|

Vn

2 logn2−p/2p

0

P|N| ≥2 logn2−p/2pdx

≤lim

ε0ε

2pb1/2−p

n

lognb

n Δn1 Δn2 Δn3 Δn4, 2.4

where

Δn1

minlogn,1/√Δn

0

P

|

Sn|

Vn

2 logn2−p/2p

P|N| ≥2 logn2−p/2pdx,

Δn2 n1/4

minlogn,1/√Δn

P

|

Sn|

Vn

2 logn2−p/2p

P|N| ≥2 logn2−p/2pdx,

Δn3 n1/2

n1/4 P

|

Sn|

Vn

2 logn2−p/2p

P|N| ≥2 logn2−p/2pdx,

Δn4

n1/2 P|Sn|

Vn

2 logn2−p/2pP|N| ≥2 logn2−p/2pdx.

2.5

Thus forΔn1, it is easy to see

Δn1≤

Δn−→0, asn−→ ∞. 2.6

Now we are in a position to estimateΔn2. From 1.6, and by applying−Xisto it, we can obtain that for large enoughnand any 0< a≤1/4, there exist C and b such thatP|Sn|/Vn>

xCe−1/2−ax2

forb < x < n1/2/b. In particular, forb < x < n1/2/b, there existsC >0 such that

P

|

Sn|

Vn > x

(6)

Hence, by Markov’s inequality and2.7, we have

Δn2≤ n1/4

minlogn,1/√Δne

2 logn2−p/2p2/4

dx

n1/4

minlogn,1/√Δn

C

2 logn2−p/2p2 dx

n1/4

minlogn,1/√Δn

ex2/4dx

n1/4

minlogn,1/√Δn

C

x2dx−→0, asn −→ ∞.

2.8

ForΔn3, by Markov’s inequality and2.7, we have

Δn3≤ n1/2

n1/4P |

Sn|

Vnn

1/4

dx

n1/2

n1/4

C

2 logn2−p/2p2 dx

e−√n/4n1/2n1/4 n1/2

n1/4

C

x2dx−→0, asn−→ ∞.

2.9

From Cauchy inequality, it follows that

|Sn|

Vn

n. 2.10

Therefore

Δn4

n1/2

P|N| ≥2 logn2−p/2pdx

n1/2

C

2 logn2−p/2p2 dx

n1/2

C

x2dx−→0, asn−→ ∞.

2.11

DenoteΔn Δn1 Δn2 Δn3 Δn4, then, since the weighted average of a sequence that converges to 0 also converges to 0, it follows that, for anyM >1,

lim

ε0ε

2pb1/2−p

n

lognb

n

E |Sn|

Vnε

2 logn2−p/2p

E

|N| −ε2 logn2−p/2p

≤lim

ε0ε

2pb1/2−p

n

lognb

n Δ

n−→0, asε0.

2.12

(7)

Proposition 2.3. Forb >−1, one has

lim

M→ ∞limε0ε

2pb1/2−p

n>Aε

lognb

n E

|N| −ε2 logn2−p/2p

0. 2.13

Proof. Note that

ε2pb1/2−p

n>Aε

lognb

n E

|N| −ε2 logn2−p/2p

ε2pb1/2−p

lognb

t

ε2 logt2−p/2pP|N| ≥xdx dt

2M

y2p/2−pb1−1

y

P|N| ≥xdx dy

2M

P|N| ≥x

x

2M

y2p/2−pb1−1dy dx

C

2M

x2p/2−pb1P|N| ≥xdx−→0, asM−→ ∞.

2.14

So this proposition is proved now.

Proposition 2.4. Forb >−1, one has

lim

M→ ∞ limε0ε

2pb1/2−p

n>Aε

lognb

n E

|Sn|

Vnε

2 logn2−p/2p

0. 2.15

Proof. Note that

ε2pb1/2−p n>Aε

lognb

n E

|Sn|

Vnε

2 logn2−p/2p

ε2pb1/2−p

n>Aε

lognb

n

0

P

|

Sn|

Vn

2 logn2−p/2p

dx

B1B2B3,

(8)

where

B1ε2pb1/2−p

n>Aε

lognb

n

n1/4

0

P

|

Sn|

Vn

2 logn2−p/2p

dx,

B2ε2pb1/2−p

n>Aε

lognb

n

n1/2

n1/4P |

Sn|

Vn

2 logn2−p/2p

dx,

B3ε2pb1/2−p

n>Aε

lognb

n

n1/2P |

Sn|

Vn

2 logn2−p/2p

dx.

2.17

ForB1, by2.7, we have

B1≤2pb1/2−p

n>Aε

lognb

n

n1/4

0

e2 logn2−p/2p2/4dx

2pb1/2−p

n>Aε

lognb

n

0

e2 logn2−p/2p2/4dx

2pb1/2−p

n>Aε

lognb

n

ε2 logn2−p/2pex2/4

dx

2pb1/2−p

lognb

t

ε2 logt2−p/2pex2/4

dxdt

C

2M

y2p/2−pb1−1

y

ex2/4dxdy

C

2M

ex2/4

x

2M

y2p/2−pb1−1dydx

C

2M

x2p/2−pb1ex2/4dx−→0, asM−→ ∞.

2.18

ForB2, using2.7again, we have

B2≤ε2pb1/2−p

n>Aε

lognb

n

n1/2n1/4P |

Sn|

Vnn

1/4ε2 logn2−p/2p

2pb1/2−p

n>Aε

lognb

n

(9)

2pb1/2−p

n>Aε

lognb

n

n1/2−n1/4e−√n/4eε22 logn2−p/p/4

2pb1/2−p

n>Aε

lognb

n e

ε22 logn2−p/p/4

2pb1/2−p

lognb

t e

ε22 logt2−p/p/4 dt

by lettingz ε

22 logt2−p/p 4

C

2M2−p/p/4z

pb1/2−p−1ezdz−→0, asM−→ ∞.

2.19

By noting that2.10, it is easily seen that

B3 0. 2.20

Combining2.18,2.19, and2.20, the proposition is proved.

Our main result follows from the propositions using the triangle inequality.

Acknowledgments

The author thanks the referees for pointing out some errors in a previous version, as well as for several comments that have led to improvements in this work. Thanks are also due to Doctor Ke-ang Fu of Zhejiang University in china for his valuable suggestion in the preparation of this paper.

References

1 J. A. Davis, “Convergence rates for probabilities of moderate deviations,” Annals of Mathematical Statistics, vol. 39, pp. 2016–2028, 1968.

2 A. Gut and A. Sp˘ataru, “Precise asymptotics in the law of the iterated logarithm,” The Annals of Probability, vol. 28, no. 4, pp. 1870–1883, 2000.

3 Y. S. Chow, “On the rate of moment convergence of sample sums and extremes,”Bulletin of the Institute of Mathematics. Academia Sinica, vol. 16, no. 3, pp. 177–201, 1988.

4 V. Bentkus and F. G ¨otze, “The Berry-Esseen bound for Student’s statistic,”The Annals of Probability, vol. 24, no. 1, pp. 491–503, 1996.

5 Q. Wang and B.-Y. Jing, “An exponential nonuniform Berry-Esseen bound for self-normalized sums,”

The Annals of Probability, vol. 27, no. 4, pp. 2068–2088, 1999.

6 E. Gin´e, F. G ¨otze, and D. M. Mason, “When is the studentt-statistic asymptotically standard normal?”

The Annals of Probability, vol. 25, no. 3, pp. 1514–1531, 1997.

(10)

8 M. Cs ¨org˝o, B. Szyszkowicz, and Q. Wang, “Darling-Erd˝os theorem for self-normalized sums,”The Annals of Probability, vol. 31, no. 2, pp. 676–692, 2003.

References

Related documents