Open Access

Limit theorems for delayed sums of random sequence

Journal of Inequalities and Applications20122012:124

https://doi.org/10.1186/1029-242X-2012-124

Received: 29 October 2011

Accepted: 31 May 2012

Published: 31 May 2012

Abstract

For a sequence of arbitrarily dependent random variables (X n )nNand Borel sets (B n )nN, on real line the strong limit theorems, represented by inequalities, i.e. the strong deviation theorems of the delayed average S n . k n ω are investigated by using the notion of asymptotic delayed log-likelihood ratio. The results obtained popularizes the methods proposed by Liu.

Mathematics Subject Classification 2000: Primary, 60F15.

Keywords

strong deviation theorem likelihood ratio delayed sums

1. Introduction

Let (a n )nNbe a sequence of real numbers and let (k n )nNbe a sequence of positive integers. The numbers
ρ n , k n = j = 1 k n a n + j - 1 / k n

are called the (forward) delayed first arithmetic means (See [1]). In [2], using the limiting behavior of delayed average, Chow found necessary and sufficient conditions for the Borel summability of i.i.d. random variables and also obtained very simple proofs of a number of well-known results such as the Hsu-Robbins-Spitzer-Katz theorem. In [3], Lai studied the analogues of the law of the iterated logarithim for delayed sums of independent random variables. Recently, Chen [4] has presented an accurate description the limiting behavior of delayed sums under a non-identically distribution setup, and has deduced Chover-type laws of the iterated logarithm for them.

Our aim in this article is to establish strong deviation theorems (limit theorem expressed by inequalities, see [5]) of delayed average for the dependent absolutely continuous random variables. By using the notion of asymptotic delayed log-likelihood ratio, we extend the analytic technique proposed by Liu [5] to the case of delayed sums. The crucial part of the proof is to construct a delayed likelihood ratio depending on a parameter, and then applies the Borel-Cantelli lemma.

Throughout, let (X n )nNbe a sequence of absolutely continuous random variables on a fixed probability space Ω , F , P with the joint density function g1, n(x1,..., x n ), n N, and f j (x), j = 1, 2,... be the the marginal density function of random variable X j . (k n )nNbe a subsequence of positive integers, such that, for every ε > 0, n = 1 exp - k n ε < .

Definition 1. The delayed likelihood ratio is defined by
L n ω = Π j = n n + k n - 1 f j X j g n , n + k n - 1 X n , . . . , X n + k n - 1 , if denominator > 0 0 , otherwise
(1.1)
Let
L ω = - lim inf n 1 k n log L n ω
(1.2)

L ω is called asymptotic delayed log-likelihood ratio, where g n , n + k n - 1 x n , , x n + k n - 1 denotes the joint density function of random vector X n , , X n + k n - 1 , ω is a sample point (with log 0 = -∞).

It will be shown in Lemma 1 that L ω 0 a.e. in any case.

Remark 1. It will be seen below that L ω has the analogous properties of the likelihood ratio in [5], Although L ω is not a proper metric among probability measures, we nevertheless consider it as a measure of "discrimination" between the dependence (their joint distribution) and independence (the product of their marginals). Obviously, L n ω = 1 , a.e. n N if (X n )nNis independent. In view of the above discussion of the asymptotic logarithmic delayed likelihood ratio, it is natural for us to think of L ω as a measure how far (the random deviation) of (X n )nNis from being independent and how dependent they are. The closer L ω approaches to 0, the smaller the deviation is.

Lemma 1. Let L n ω be define as above, then
lim sup n 1 k n log L n ω 0 , a . e .
(1.3)
Proof. Let B = x n , , x n + k n - 1 : g n , n + k n - 1 x n , , x n + k n - 1 > 0 . Since
E L n ω = x n , , x n + k n - 1 B j = n n + k n - 1 f j x j g n , n + k n - 1 x n , , x n + k n - 1 . g n , n + k n - 1 x n , , x n + k n - 1 d x n d x n + k n - 1 = x n , , x n + k n - 1 B j = n n + k n - 1 f j x j d x n d x n + k n - 1 x n , , x n + k n - 1 R k n j = n n + k n - 1 f j x j d x n d x n + k n - 1 = 1 .
From Markov inequality, for every ε > 0, we have
P 1 k n log L n ω ε = P L n ω exp k n ε 1 exp - k n ε .
Hence
n = 1 P 1 k n log L n ω ε n = 1 exp - k n ε < .
By Borel-Cantelli lemma, we have
P lim sup n 1 k n log L n ω 2 ε = 0 ,

for any ε > 0, (1.3) follows immediately. □

2. Main results and proofs

Theorem 1. Let (X n )nN, L n ω , L ω be defined as above, (B n )nNbe a sequence of Borel sets of the real line. Let S n , k n ω = 1 k n j = n n + k n - 1 1 B j X j , and assume
c = lim sup n 1 k n j = n n + k n - 1 P X j B j ,
(2.1)
then
lim sup n S n , k n ω L ω + c 2 , a . e .
(2.2)

where 1 B n ( ) be the indicator function of B n .

Proof. Assume s > 0 to be a constant, and let
h j x j = s 1 B j x j f j x j 1 + s - 1 B j f j x j d x j , j = 1 , 2 ,
(2.3)
It is not difficult to see that h j x j d x j = 1 , j = 1, 2,... Let
Λ n s , ω = Π j = n n + k n - 1 h j X j g n , n + k n - 1 X n , . . . , X n + k n - 1 , if denominator > 0 0 , otherwise
(2.4)
From Lemma 1, there exists A s F , P(A(s)) = 1, such that
lim sup n 1 k n log Λ n s , ω 0 , ω A s
(2.5)
Since B j f j x j d x j = P X j B j , by (2.3) we have
j = n n + k n - 1 h j x j = j = n n + k n - 1 s 1 B j x j f j x j 1 + s - 1 B j f j x j d x j = s j = n n + k n - 1 1 B j x j j = n n + k n - 1 f j x j 1 + s - 1 P X j B j
(2.6)
It follows from (1.1), (2.4) and (2.6) that
log Λ n s , ω = j = n n + k n - 1 1 B j X j log s - j = n n + k n - 1 log 1 + s - 1 P X j B j + log L n ω
(2.7)
(2.5) and (2.7) yield
lim sup n 1 k n log s j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j + log L n ω 0 , ω A s
(2.8)
Let s > 1, dividing the two sides of (2.8) by log s, we have
lim sup n 1 k n j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j log s + log L n ω log s 0 , ω A s
(2.9)
By (1.2), (2.9) and the property lim sup n (a n - b n ) ≤ d lim sup n (a n - c n ) ≤ lim sup n (b n - c n ) + d, one gets
lim sup n 1 k n j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j log s L ω log s , ω A s
(2.10)
By (2.10) and the property of the superior above and the inequality 0 < log(1+x) ≤ x(x > 0), we obtain
lim sup n S n , k n ω lim sup n 1 k n j = n n + k n - 1 log 1 + s - 1 P X j B j log s + L ω log s lim sup n 1 k n j = n n + k n - 1 s - 1 P X j B j log s + L ω log s c s - 1 log s + L ω log s , ω A s
(2.11)
(2.11) and the inequality 1 - 1 s < log s s > 1 imply
lim sup n S n , k n ω c s + s L ω s - 1 , ω A s
(2.12)
Let D be a set of countable real numbers dense in the interval (1, +∞), and let A* = ∩sDA(s), g(s, x) = cs + sx/(s - 1), then we have by (2.12)
lim sup n S n , k n ω g s , L ω , ω A * , s D
(2.13)
Let c > 0, it easy to see that if L ω > 0 , a.e., then, for fixed ω, g s , L ω as a function of s attains its smallest value g 1 + L ω / c , L ω = 2 c L ω + L ω + c on the interval (1, +∞), and g(s, 0) is increasing on the interval (1, +∞) and lims→1+ g(s, 0) = 0. For each ω A* ∩ A(1), if L ω , take κ n (ω) D, n = 1, 2,..., such that κ n ω 1 + L ω / c . We have by the continuity of g s , L ω with respect to s,
lim n + g κ n ω , L ω = L ω + c 2 ,
(2.14)
By (2.13), we obtain
lim sup n S n , k n g κ n ω , L ω , n = 1 , 2 ,
(2.15)
(2.14) and (2.15) imply
lim sup n S n , k n ω L ω + c 2 , ω A * A 1
(2.16)

If L ω = , (2.16) holds trivially. Since P (A* ∩ A(1)) = 1, (2.2) holds by (2.16), when c > 0.

When c = 0, we have by letting s = e in (2.11),
lim sup n S n , k n ω L ω , ω A e
(2.17)

since P (A(e)) = 1, (2.2) also holds by (2.17) when c = 0. □

Theorem 2. Let (X n )nN, L n ω , L ω , (B n )nN, S n , k n ω be defined as in Theorem 1 and assume
c = lim inf n 1 k n j = n n + k n - 1 P X j B j ,
(2.18)
then, if 0 L ω c a.e., then
lim inf n S n , k n ω c - 2 c L ω , a . e .
(2.19)
Proof. Let 0 <s < 1, dividing the two sides of (2.8) by log s, we have
lim inf n 1 k n j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j log s + log L n ω log s 0 , ω A s
(2.20)
By (1.2), (2.20) and the property lim inf n (a n - b n ) ≥ d lim inf n (a n - c n ) ≥ lim inf n (b n - c n ) + d, one gets
lim inf n 1 k n j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j log s L ω log s , ω A s
(2.21)
By (2.21) and the property of the inferior above and the inequality log(1+ x) ≤ x(-1 <x ≤ 0), we obtain
lim inf n S n , k n ω lim inf n 1 k n j = n n + k n - 1 log 1 + s - 1 P X j B j log s + L ω log s lim n inf 1 k n j = n n + k n - 1 s - 1 P X j B j log s + L ω log s c s - 1 log s + L ω log s , ω A s
(2.22)
(2.22) and the inequality 1 - 1 s < log s and log s < s - 1 (0 <s < 1) imply
lim inf n S n , k n ω c s + L ω s - 1 , ω A s A 1
(2.23)
Let D' be a set of countable real numbers dense in the interval (0, 1), and let A * = s D A s , h(s, x) = c's + x/(s - 1), then we have by (2.23)
lim inf n S n , k n ω h s , L ω , ω A * , s D
(2.24)
Let c' > 0, it easy to see that if 0 < L ω < c , a.e., then, for fixed ω, h s , L ω as a function of s attains its maximum value h 1 - L ω / c , L ω = c - 2 c L ω , on the interval (0, 1), and h(s, 0) is increasing on the interval (0, 1) and lims→1+ h(s, 0) = c'. For each ω A*A(1), if L ω , take l n (ω) D', n = 1, 2,..., such that l n ω 1 - L ω / c . We have by the continuity of h s , L ω with respect to s,
lim n + h l n ω , L ω = c - 2 c L ω ,
(2.25)
By (2.24), we obtain
lim inf n S n , k n h l n ω , L ω , n = 1 , 2 ,
(2.26)
(2.25) and (2.26) imply
lim inf n S n , k n ω c - 2 c L ω , ω A * A 1
(2.27)

If L ω = , (2.27) holds trivially. Since P (A*A(1)) = 1, (2.19) holds by (2.27), when c' > 0. (2.19) also holds trivially when c' = 0. □

Remark 2. In case L ω > c 0 , a.e., we cannot get a better lower bound of lim inf n S n , k n ω . This motivates the following problem: under the conditions of Theorem 2, how to get a better lower bound of lim inf n S n , k n ω in case of L ω > c 0 , a.e.?

Definition 2. (Generalized empirical distribution function) Let (X n )nNbe identically distribution with common distribution function F , for each m, n N, let
F m , n x = 1 n k = m m + n - 1 1 X k x .

F m, n = the observed frequency of values that are ≤ x from time m to m + n - 1. The F1,nis the usual empirical distribution function, hence the name given above.

In particular, let B = (-∞, x], x R in Theorems 1 and 2, we can get a strong limit theorem for the generalized empirical distribution function.

Corollary 1. Let (X n )nNbe i.i.d. random variables with common distribution function F, let B n = (-∞, x], n = 1, 2,..., then
lim n F n , n + k n - 1 x = F x , a . e .
Corollary 2. Let (X n )nNbe independent random variables and (B n )nNbe as Theorem 1, then
lim n 1 k n j = n n + k n - 1 1 B j X j - P X j B j = 0 , a . e .
(2.28)
Proof. Note that P (X j B j ) ≤ 1, j = 1, 2,... and in this case, 0 ≤ c, c' ≤ 1, L ω = 0 a.e., we have by (2.11)
lim sup n 1 k n j = n n + k n - 1 1 B j X j - j = n n + k n - 1 log 1 + s - 1 P X j B j log s 0 , ω A s
(2.29)
by (2.29) and the property of the superior above and the inequality 0 ≤ log(1+x) ≤ x(x > 0), we obtain
lim sup n 1 k n j = n n + k n - 1 1 B j X j - P X j B j lim sup n 1 k n j = n n + k n - 1 log 1 + s - 1 P X j B j log s - P X j B j lim sup n 1 k n j = n n + k n - 1 s - 1 P X j B j log s - P X j B j s - 1 log s - 1 , ω A s
(2.30)
Analogously as in the proof of Theorem 1, we obtain
lim sup n 1 k n j = n n + k n - 1 1 B j X j - P X j B j 0 , a . e .
(2.31)

Similarly, we have lim inf n 1 k n j = n n + k n - 1 1 B j X j - P X j B j 0 , a.e. hence (2.28) follows immediately. □

Remark 3. Let B n = B, Corollary 2 implies that lim S n , k n k n = P X 1 B which gives the strong law of large numbers for the delayed arithmatic means.

Declarations

Acknowledgements

This work is supported by The National Natural Science Foundation of China (Grant No. 11071104) and the An Hui University of Technology research grant: D2011025. The authors would like to thank two referees for their insightful comments which resulted in improving Theorems 1, 2 and Corollary 2 significantly.

Authors’ Affiliations

(1)
Department of Mathematics & Physics, HeFei University
(2)
Faculty of Mathematics & Physics, AnHui University of Technology

References

  1. Zygmund A: Trigonometric Series 1. Cambridge Universitiy Press, Cambridge; 1959.Google Scholar
  2. Chow YS: Delayed sums and Borel summability for independent, identically distributed random variables. Bull Inst Math Academia Sinica 1972, 1: 207–220.Google Scholar
  3. Lai TL: Limit theorems for delayed sums. Ann Probab 1974, 2(3):432–440. 10.1214/aop/1176996658View ArticleGoogle Scholar
  4. Chen PY: Limiting behavior of delayed sums under a non-identically distribution setup. Ann Braz Acad Sci 2008, 80(4):617–625.View ArticleGoogle Scholar
  5. Liu W: Strong deviation theorems and analytical method. Academic press, Beijing; 2003.Google Scholar

Copyright

© Fang-qing and Zhong-zhi; licensee Springer. 2012

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.