Abstract
We consider the asymptotic behavior of the convolution \({P}^{{_\ast}n}(\sqrt{n}A)\) of a k-dimensional probability distribution P(A) as \(n \rightarrow \infty \) for A from the σ-algebra \({\mathfrak{M}}^{k}\) of Borel subsets of Euclidian space R k or from its subclasses (often appearing in mathematical statistics). We will deal with two questions: construction of asymptotic expansions and estimating the remainder terms by using necessary and sufficient conditions. The most widely and deeply investigated cases are those where \({P}^{{_\ast}n}(\sqrt{n}A)\) are approximated by the k-dimensional normal laws \({\Phi }^{{_\ast}n}(A\sqrt{n})\) or by the accompanying ones \({\mathit{e}}^{n(P-E_{0})}\). In this and other papers, estimating the remainder terms, we extensively use the method developed in the candidate thesis of Yu. V. Prokhorov (Limit theorems for sums of independent random variables. Candidate Thesis, Moscow, 1952) (adviser A. N. Kolmogorov) and there obtained necessary and sufficient conditions (see also Prokhorov (Dokl Akad Nauk SSSR 83(6):797–800 (1952) (in Russian); 105:645–647, 1955 (in Russian)).
Mathematics Subject Classification (2010): 60F05, 60F99
Access provided by Autonomous University of Puebla. Download conference paper PDF
Similar content being viewed by others
Keywords
- Probability distributions in R k
- Convolutions
- Bergström identity
- Appell polynomials
- Chebyshev–Cramer asymptotic expansion
1 Introduction
We first present three theorems from the thesis of Yu. V. Prokhorov [10]. Let
be a sequence of independent identically distributed random variables with distribution function \(F(x) = P\{\xi _{1} < x\}\).
Theorem P4.
Let F(x) satisfy one of the following two conditions:
-
1.
F(x) is a discrete distribution function;
-
2.
There exists an integer n 0 such that \({F}^{{_\ast}n_{0}}(x)\) has an absolutely continuous component.
Then there exists a sequence \(\{G_{n}(x)\}\) of infinitely divisible distribution functions such that
where \(\Vert \cdot \Vert\) stands for the total variation.
Theorem P5.
In order that
for some appropriately chosen constants B n > 0 and A n and a stable distribution function G(x), the following conditions are necessary and sufficient:
-
1.
\({F}^{{_\ast}n}(xB_{n} + A_{n}) \rightarrow G(x),\quad n \rightarrow \infty ,\quad x \in {R}^{1}\);
-
2.
There exists n 0 such that
$$\int\limits_{-\infty }^{\infty }p_{ n_{0}}(x)\,dx > 0,$$
where \(p_{n_{0}}(x) = \frac{d} {dx}F_{(x)}^{{_\ast}n_{o}}\) .
Theorem P6.
Suppose that \(\xi _{1}\) takes only the values \(m = 0,\pm 1,\ldots \) and that the stable distribution function G(x) has a density g(x). Then
if and only if the following two conditions are satisfied:
-
1.
\({F}^{{_\ast}n}(xB_{n} + A_{n}) \rightarrow G(x),\quad n \rightarrow \infty ,\quad x \in {R}^{1}\);
-
2.
The maximal step of the distribution of ξ 1 equals 1.
In the case where \(G(x) = \Phi (x)\) is the standard Gaussian distribution function, the following statement is proved.
Theorem 1.
Let \(\xi _{1}\) have 0 mean and unit variance. In order that
for some \(\delta \in (0,1]\) , the following two conditions are necessary and sufficient:
-
1.
\(\sup _{x}\big{\vert }{F}^{{_\ast}n}(x\sqrt{n}) - \Phi (x)\big{\vert } = O({n}^{-\delta /2}),\quad n \rightarrow \infty ;\)
-
2.
There exists n 0 such that the distribution function \({F}^{{_\ast}n_{0}}(x)\) has an absolutely continuous component.
The theorem is proved in [2]. In the same paper, a sequence of random variables \(\xi _{1},\ldots ,\xi _{n},\ldots \) with values \(m = 0,\pm 1,\pm 2,\ldots \) is also considered. In this case, the following statement is proved.
Theorem 2.
In order that
for some \(\delta \in (0,1]\) , the following two conditions are necessary and sufficient :
-
1.
\(\sup _{x}\big{\vert }{F}^{{_\ast}n}(x\sqrt{n}) - \Phi (x)\big{\vert } = O({n}^{-\delta /2}),\quad n \rightarrow \infty \) ;
-
2.
The maximal step of the distribution of ξ 1 is 1.
In the case where P(A) is a probability distribution defined in the k-dimensional space R k, and \(\Phi (A)\) is the standard k-dimensional normal distribution, the following theorem is proved in [3].
Theorem 3.
In order that
the following two conditions are necessary and sufficient :
-
1.
\(\sup _{\Vert \vec{t}\Vert =1}\sup _{x\in {R}^{1}}\big{\vert }{P}^{{_\ast}n}(\sqrt{n}A_{x}(\vec{t}))) - \Phi (A_{x}(\vec{t}))\big{\vert } = O({n}^{-\delta /2})\quad \mbox{ as}\quad n \rightarrow \infty ,\) where \(A_{x}(\vec{t}) =\{\vec{ u} :\, (\vec{t},\vec{u}) < x\}\) , \(\Vert \vec{t}\Vert\) is the length of a vector \(\vec{t} \in {R}^{k}\) , and \((\vec{u},\vec{t})\) denotes the inner product in R k ;
-
2.
There exists n 0 such that the distribution function \({F}^{{_\ast}n_{0}}\) has a absolutely continuous component.
The statements of Theorems 1–3 remain valid if one replaces \(\Phi (A)\) by “long” Chebyshev–Cramer asymptotic expansions with appropriate changes in condition (1) and with no changes in the Prokhorov conditions (in the theorems, conditions (2)); see [3].
2 Appell Polynomials
Recall that a sequence of polynomials \(g_{n}(x)\), n = 1, 2, …, is called an Appell polynomial set if
see [6], p. 242.
Often, by Appell polynomials are meant the polynomials
defined by
for \(\vert z\vert < \tau \) (see [5, 8]).
The coefficients \(q_{j\,l}\) satisfy the recursion formula
for \(j = 1,2,\ldots \), \(l = 1,2,\ldots ,j - 2\) (see [8]). For l < 0, q jl = 0, and
It is known [8] that
for \(j = 1,2,\ldots \), \(l = 0,1,\ldots ,j - 1\).
Estimating the remainder terms of asymptotic expansions, we will use the following lemma.
Lemma 1.
We have
The lemma can be proved by induction using (3).
Let us now estimate the remainder term
Here z may be a complex number, e.g., the difference of characteristic functions of random vectors, \(\tau > 0\), and \(\vert z\vert < \tau \); \(s = 1,2,\ldots \).
Lemma 2.
We have
Proof.
From (2) and (3) it follows that
Now it remains to apply inequality (4), and the lemma follows after a simple calculation.
3 Expansion of Convolutions of Measures by Appell Polynomials
Consider the convolutions of generalized finite-variation measures μ(B), \(B \in {\mathfrak{M}}^{k}\):
where μ0 is the Dirac measure, \(\vec{0} = (0,0,\ldots ,0) \in {R}^{k}\), \(n = 1,2,\ldots \);
It is obvious that
Theorem 4.
If \(\Vert \mu \Vert < n\) , we have the asymptotic expansion
where \(n = 1,2,\ldots \) , and
is the Appell polynomial with the powers of μ are understood in the convolution sense.
Proof.
Obviously,
where
and \(C_{\nu }^{(j)}\) is the Stirling number of the first kind.
From the last two equalities it follows that
Since \(C_{\nu }^{(0)} = 1\) and
we obtain
The theorem is proved.
Theorem 5.
Let μ and \(\mu _{1}\) be generalized finite-variation measures in R k . Then, for every Borel set \(B \in {\mathfrak{M}}^{k}\) , we have
where \(n = 1,2,\ldots \) , and
Proof.
When \(\Vert \mu \Vert < n\), the remainder term is
From this it follows that
Here,
From this and from (5) the theorem follows.
Suppose that the probability distribution has an inverse generalized measure \({G}^{-{_\ast}}\), i.e.,
where E 0 is the degenerate \(k\)-dimensional measure concentrated at \(\vec{0} \in {R}^{k}\). Such a property is possessed by accompanying probability distributions \({\mathit{e}}^{F-E_{0}}\), i.e., \({G}^{-{_\ast}}\ =\ { \mathit{e}}^{-(F-E_{0})}\).
Theorem 6.
Let F be a k-dimensional probability distribution, let a probability distribution G have an inverse \({G}^{-{_\ast}}\) , and let \(\varrho =\Vert (F - G) {_\ast} {G}^{-{_\ast}}\Vert < 1\) . Then
where
To estimate the remainder term
we use
and
It is obvious that
where
From (7) and (8) there follows an estimate of the remainder term in the asymptotic expansion (6).
4 Expansion of a Convolution by Accompanying Probability Measures
Every k-dimensional probability measure P satisfies the identity
where
with \(P\{\xi _{1} = m\} = \frac{m + 1} {(m + 2)!}\), \(m = 0,1,2,\ldots \).
From (9) and (10) it follows that
for l = 1, 2, …, where \(z_{l} = \xi _{1} + \xi _{2} + \cdots + \xi _{l}\) -is the sum of i.i.d. random variables \(\xi _{1},\xi _{2},\ldots ,\xi _{l}\).
It is obvious that, for all P,
If \(\Vert {(P - E_{0})}^{{_\ast}2}\Vert < \frac{4} {1+{\mathit{e}}^{2}}\), then
and for the convolution \({P}^{{_\ast}n}\), we can apply Theorem 6:
where
and
Let us estimate the remainder term
From (11) it follows that
and
Theorem 7.
Suppose that \(\Vert {(P - E_{0})}^{{_\ast}2}\Vert < \frac{4} {1+{\mathit{e}}^{2}}\) . Then, for all Borel sets \(B \in {\mathfrak{M}}^{k}\) ,
where
\(\varrho =\Vert \mu \Vert\) , and
5 Asymptotic Bergström Expansion
For any \(k\)-dimensional probability distributions P and Q,
(the Bergström identity). Here, for \(s + 1 < n\),
Let \(\Theta \) be a negative hypergeometric random variable taking the natural values \(m = s + 1,s + 2,\ldots ,n\) with probabilities
Then we can rewrite the remainder term as
where
Lemma 3.
Suppose that P and Q have finite jth-order absolute moments and that
for \(r = 1,2,\ldots ,j\) and \(\vec{t} \in {R}^{k}\) . Then
for \(l = 0,1,\ldots ,(j + 1)m - 1\) and \(\vec{t} \in {R}^{k}\) .
Remark.
If the first moments of P and Q coincide, then
for \(l = 0,1,\ldots ,3m - 1\).
The lemma is proved by using characteristic functions and the Faa de Bruno formula that can be found, e.g., in [8].
Since
where \(C_{\nu }^{(j)}\) is the Stirling number of the first kind, \(C_{\nu }^{(0)} = 1\), and
we have, for \(1 \leq s < n\),
Now, the Bergström identity becomes
Let us now consider the cases where Q(B) is the normal k-dimensional distribution \(\Phi (B) = P\{\boldsymbol \xi \in B\}\), \(\boldsymbol \xi \sim N_{k}(\vec{0},\Sigma )\), where Σ is a nondegenerate matrix of second moments.
Suppose that the expectation vectors and second-moment matrices of P(B) and \(\Phi (B)\) coincide.
Theorem 8.
Suppose that the probability distribution \(P(B) = P\{\boldsymbol \eta \in B\}\) has finite absolute moments of order \(2 + \delta \) with \(0 < \delta \leq 1\) . Then there exists a constant C, depending only on k, s, and δ, such that
for \(1 \leq \nu \leq s\) , where
and \(\boldsymbol {\eta }^{T}\) is the transpose of the vector \(\boldsymbol \eta \) .
Theorem 8 is proved in [4]. H. Bergström proved that (see [1])
We will estimate the remainder term \(r_{n}^{(s+1)}(B)\) for all convex Borel sets \(B \in {\mathfrak{N}}^{k}\).
Theorem 9.
Suppose that the assumptions of Theorem 8 are satisfied and that the characteristic function of the random vector \(\boldsymbol \eta _{1}\) satisfies Cramer condition (C):
Then
The theorem is proved in [4].
In the one-dimensional case (k = 1), Bergström [1] proved that from his asymptotic expansion there follows the Chebyshev–Cramer asymptotic expansion.
For k > 1 and \(Q(B) = \Phi (B)\), from (13) it follows that
The formal asymptotic expansion of the density \(p_{\nu }(\vec{y})\) of the convolution
is
for \(1 \leq \nu \leq s\), where \(\vert \Sigma \vert \) denotes the determinant of the matrix Σ.
Let \(\boldsymbol \xi _{\epsilon } \sim N_{k}(\vec{0},(1 + \epsilon )\Sigma ))\) be a k-dimensional normal random vector. If \(\epsilon = 0\), then
From (14) and (15) we get the following formal expansion of the convolution \({P}^{{_\ast}n}(B\sqrt{n})\):
where
The formal expansions are obtained by means of the characteristic functions.
6 Expansion of a Convolution by χ2-Distributions
Let \(\boldsymbol \xi _{\mu } \sim N_{k}(\boldsymbol \mu ,\Sigma )\) be a normal k-dimensional random vector with nondegenerate covariation? matrix Σ. The random variable
has the \({\chi }^{2}\)-distribution with k degrees of freedom, and the random variable
has the noncentral χ2-distribution with k degrees of freedom and noncentrality parameter
The distribution function of \(\chi _{k}^{2}(\delta )\) is
Let
be the sum of i.i.d. k-dimensional vectors \(\boldsymbol \eta _{1},\ldots ,\boldsymbol \eta _{n},\ldots \) with zero mean vector \(\vec{0}\ \in \ {R}^{k}\) and nondegenerate covariation matrix \(\Sigma \). Let \(\boldsymbol \xi \sim N_{k}(\vec{0},\Sigma )\) and
We are interested in an asymptotic expansion of
where \(P(\sqrt{n}A_{x}) = P\big\{ \frac{\boldsymbol \eta _{1}} {\sqrt{n}} \in A_{x}\big\}\), i.e., the difference
for x > 0.
Denote by \(\widehat{P}(\vec{t})\) and \(\widehat{\Phi }(\vec{t})\) the characteristic functions of the vectors \(\boldsymbol \eta _{1}\) and \(\boldsymbol \xi \). From the Bergström identity (13) it follows that
where
The Fourier transform is
By the change of variables \(\vec{v} = \sqrt{\frac{n - \nu } {n}} \vec{t}\) we obtain
where
From (17) to (20) after the change of variables \(\vec{u} =\vec{ y}\sqrt{ \frac{n} {n-\mu }} - \frac{\vec{x}} {\sqrt{n-\nu }}\) it follows that
where
is the noncentrality parameter of the \(\chi _{k}^{2}(\delta (\vec{x}))\)-distribution. From (16) it follows that
Now the asymptotic Bergström expansion writes as
To estimate the remainder term, we applied Theorem 9. Thus, we have proved the following:
Theorem 10.
Suppose that a random vector \(\boldsymbol \eta _{1}\) has a finite absolute moment of order \(2 + \delta \) for some \(0 < \delta \leq 1\) and that the characteristic function \(\widehat{P}(\vec{t})\) satisfies the Cramer condition
Then
for all x > 0 and s = 1,2,….
If instead of considering the χ k 2 random variable, we change t, F, μ, etc., then we have also to change the definition of the set A x and to replace the Cramer condition (C) by the Prokhorov [11, 12] condition that there exists n 0 such that the convolution \({P}^{{_\ast}n_{0}}(\vec{x})\) has an absolutely continuous component.
References
Bergström, H.: On asymptotic expansions of probability functions. Skand. Aktuartiedskr. H. 34 (1–2), 1–34 (1951)
Bikelis, A.: On the accuracy of Gaussian approximation of the distribution of a sum of independent identically distributed random variables. Liet. mat. rink. 11(2), 237–240 (1971) (in Russian)
Bikelis, A.: On the central limit theorem in R k. I–III. Liet. mat. rink. 11(1), 27–58 (1971); 12(1), 3–84 (1972); 12(3), 19–35 (1972) (in Russian)
Bikelis, A., Mogyorodi, J.: On asymptotic expansion of the convolution of n multidimensional distribution functions. Liet. mat. rink. 11(3), 433–443 (1970) (in Russian)
Dickey, J.M.: Expansions of t densities and related complete integrals. Ann. Math. Stat. 38(2), 503–505 (1967)
Erdélyi, A.: Higher Transcendental Functions, vol. 3. McGraw-Hill, New York (1953)
Fisher, R.A.: The general sampling distribution of the multiple correlation coefficient. Proc. R. Soc. London. 121A, 654–673 (1928)
Kalinin, V.M.: Special functions and limit properties of probability distributions. Notes Sci. Sem. LOMI. 13, 5–137 (1968) (in Russian)
Patnaik, P.B.: The non-central χ2-and F-distributions and their applications. Biometrika 36, 202–232 (1949)
Prokhorov, Yu.V.: Limit theorems for sums of independent random variables. Candidate Thesis, Moscow (1952) (in Russian)
Prokhorov, Yu.V.: A local theorem for densities. Dokl. Akad. Nauk SSSR 83(6), 797–800 (1952) (in Russian)
Prokhorov, Yu.V.: On sums of identically distributed random variables. Dokl. Akad. Nauk SSSR 105, 645–647 (1955) (in Russian)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2013 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Bikelis, A. (2013). Asymptotic Expansions for Distributions of Sums of Independent Random Vectors. In: Shiryaev, A., Varadhan, S., Presman, E. (eds) Prokhorov and Contemporary Probability Theory. Springer Proceedings in Mathematics & Statistics, vol 33. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-33549-5_6
Download citation
DOI: https://doi.org/10.1007/978-3-642-33549-5_6
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-642-33548-8
Online ISBN: 978-3-642-33549-5
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)