×

An inequality for the Lévy distance between two distribution functions and its applications. (Ukrainian, English) Zbl 1224.60170

Teor. Jmovirn. Mat. Stat. 81, 53-64 (2009); translation in Theory Probab. Math. Stat. 81, 59-70 (2010).
The authors deal with the Lévy distance \({\mathcal L}(F,G)\) between two distribution functions \(F\) and \(G\) defined as \({\mathcal L}(F,G)=\inf\{h: G(x-h)-h\leq F(x)\leq G(x+h)+h\text{ holds for all }x\}\). The Lévy distance is much less popular in probability theory than the uniform distance \(\Delta(F,G)=\sup_{x\in\mathbb R}| F(x)-G(x)| \). Note that always \({\mathcal L}(F,G)\leq\Delta(F,G)\). The advantage of the Lévy distance becomes evident if one considers the weak convergence \(F_n\overset {w}{} G\) as \(n\to\infty\), which is equivalent to \({\mathcal L}(F_n,G)\to0\) as \(n\to\infty\) (see, for example, [B. V. Gnedenko and A. N. Kolmogorov, Limit distributions for sums of independent random variables. Cambridge: Addison-Wesley Publishing Company (1954; Zbl 0056.36001)]). If \(G\) is continuous, then the weak convergence \(F_n\overset {w}{} G\) as \(n\to\infty\) is equivalent to \(\Delta(F_n,G)\to0\) as \(n\to\infty\); however, this property may fail if \(G\) has discontinuities. The authors first prove the following theorem which is a generalization of the corresponding result published by K.-H. Indlekofer and O. I. Klesov [Int. J. Pure Appl. Math. 47, No. 2, 235–241 (2008; Zbl 1165.60013)].
Theorem 1. Let \(F\) be an arbitrary distribution function and let \(\Phi\) be the distribution function of the standard \(N(0, 1)\) Gaussian law. Denote by \(L={\mathcal L}(F,\Phi)\) the Lévy distance between \(F\) and \(\Phi\). Suppose the moment of order \(p > 0\) exists for \(F\). Then there exists a function \(g(x)\) defined on \((0, 1)\), depending only on \(p\), and such that \(\lim_{s\to+0}g(s)=0\) and \[ | F(x)-\Phi(x)| \leq\frac{\lambda_p+g(L)}{1+| x| ^p},\quad x\in\mathbb R, \] if \(0\leq L\leq 1\).
Next, the authors deal with the generalized global version of the central limit theorem proved by R. P. Agnew [Proc. Natl. Acad. Sci. USA 40, 800–804 (1954; Zbl 0055.36703)].
Theorem C. Let \(\{F_n\}\) be a sequence of distribution functions such that \(\int_{-\infty}^{+\infty}xdF_n(x)=0\), \(\int_{-\infty}^{+\infty}x^2dF_n(x)=1\). Let \(F_n\overset {w}{} \Phi\) as \(n\to\infty\). Then, for all \(r>1/2\), \[ \int_{-\infty}^{+\infty}| F_n(x)-\Phi(x)| ^{r}\,dx\to0,\quad n\to\infty. \] The authors complete Agnew’s result by considering the case \(r = 1/2\).
Theorem 3. Let \(\{F_n\}\) be a sequence of distribution functions satisfying the conditions of the previous theorem. Then \[ \int_{-\infty}^{+\infty} \frac{| F_n(x)-\Phi(x)| ^{1/2}} {(\ln(1 + | x| ))^{1+\delta}}\,dx\to 0,\quad n\to\infty \] for all \(\delta>0\).
A version of the results of G. Laube [Metrika 20, 103–105 (1973; Zbl 0257.60013)] and A. Rosalsky [Int. J. Math. Math. Sci. 11, No. 2, 365–374 (1988; Zbl 0645.60032)] on the global version of the central limit theorem is proved for the case \(r = 1/2\).
Theorems which generalize the corresponding results by P. L. Hsu and H. Robbins [Proc. Natl. Acad. Sci. USA 33, 25–31 (1947; Zbl 0030.20101)] and C. C. Heyde [J. Appl. Probab. 12, 173–175 (1975; Zbl 0305.60008)] on the complete convergence of a sequence of random variables are proved.

MSC:

60J05 Discrete-time Markov processes on general state spaces
PDFBibTeX XMLCite
Full Text: DOI