Refined Self-normalized Large Deviations for Independent Random Variables

Let X ₁,X ₂,…?, be independent random variables with EX _i =0 and write \(S_{n}=\sum_{i=1}^{n}X_{i}\) and \(V_{n}^{2}=\sum_{i=1}^{n}X_{i}^{2}\). This paper provides new refined results on the Cramér-type large deviation for the so-called self-normalized sum S _n /V _n . The major techniques used to derive these new findings are different from those used previously.     

A nonclassical law of the iterated logarithm for self-normalized partial sums

Let {X _n ,?n≧1} be a sequence of nondegenerate, symmetric, i.i.d. random variables which are in the domain of attraction of the normal?law?with zero means and possibly infinite variances. Denote ${S_{n}=\sum_{i=1}^{n} X_{i}}$ , ${V_{n}^{2}=\sum_{i=1}^{n} X_{i}^{2}}$ . Then we prove that there is a sequence of positive constants {b(n),?n≧1} which is defined by Klesov and Rosalsky [11], is monotonically approaching infinity and is not asymptotically equivalent to loglogn but is such that $\displaystyle \limsup_{n\to\infty} \frac{|S_n|}{\sqrt{2V_n^2b(n)}}= 1$ almost surely if some additional technical assumptions are imposed.     

Limiting distribution of sums of nonnegative stationary random variables

Summary Let {X _n,j,−∞<j<∞∼,n≧1, be a sequence of stationary sequences on some probability space, with nonnegative random variables. Under appropriate mixing conditions, it is shown thatS _n=X_n,1+…+X _n,n has a limiting distribution of a general infinitely divisible form. The result is applied to sequences of functions {f _n(x)∼ defined on a stationary sequence {X _j∼, whereX _n.f=f_n(X_j). The results are illustrated by applications to Gaussian processes, Markov processes and some autoregressive processes of a general type. This paper represents results obtained at the Courant Institute of Mathematical Sciences, New York University, under the sponsorship of the National Sciences Foundation, Grant MCS 82-01119.     

An ODE’s Version of the Formula of Baker, Campbell, Dynkin and Hausdorff and the Construction of Lie Groups with Prescribed Lie Algebra

If ${\mathcal{L} = {\sum_{j=1}^m} {X_j^2} + X_0}If L = ?_j=1^m X_j² + X₀{\mathcal{L} = {\sum_{j=1}^m} {X_j^2} + X_0} is a H?rmander partial differential operator in \mathbbR^N{\mathbb{R}^N}, we give sufficient conditions on the vector fields X _j ’s for the existence of a Lie group structure \mathbbG = (\mathbbR^N, *){\mathbb{G} = (\mathbb{R}^N, *)} (and we exhibit its construction), not necessarily nilpotent nor homogeneous, such that L{\mathcal{L}} is left invariant on \mathbbG{\mathbb{G}}. The main tool is a formula of Baker-Campbell-Dynkin-Hausdorff type for the ODE’s naturally related to the system of vector fields {X ₀, . . . , X _m }. We provide a direct proof of this formula in the ODE’s context (which seems to be missing in literature), without invoking any result of Lie group theory, nor the abstract algebraic machinery usually involved in formulas of Baker-Campbell-Dynkin-Hausdorff type. Examples of operators to which our results apply are also furnished.     

On a certain exponential inequality for Gaussian processes

If X=(X _j ) _j=1 ^m is a zero-mean Gaussian stochastic process and $\sigma _{j}=\left( E{\big[} X_{j}^{2}{\big]} \right) ^{1/2},$ j=1,...,m, Tsirel’son (Theory Probab. Appl., 30, 820–828, 1985) and more explicitly Vitale (Ann. Probab., 24, 2172–2178, 1996 and A log-concavity proof for a Gaussian exponential bound. In: Hill, T.P., Houdré, C. (eds.) Advances in Stochastic Inequalities, Contemporary Mathematics, vol. 234, pp. 209–212. AMS, Providence, RI, 1999) applied results from Brunn–Minkowski theory to show that X satisfies the following inequality: $$ E\left[ \exp \left( \max_{1\leq j\leq m}{\bigg(}X_{j}-\frac{\sigma _{j}^{2}}{2} {\bigg)}\right) \right] \leq \exp \left( E\left[ \max_{1\leq j\leq m}X_{j}\right] \right). $$ In this paper a more general inequality will be derived using a known formula for Gaussian integrals. In particular, it also follows that $$ {\small \ }E\left[ \exp \left( \min_{1\leq j\leq m}{\bigg(}X_{j}-\frac{\sigma _{j}^{2}}{2}{\bigg)}\right) \right] \leq \exp \left( E\left[ \min_{1\leq j\leq m}X_{j}\right] \right) . $$ In the last section of this article the above exponential inequalities are combined with a well known variant of the Slepian lemma to compare certain option prices in the Black–Scholes and Bachelier models.     

Distribution and moment convergence of martingales

Summary If where {X _{n j} ,ℱ _{n j} 1≦j≦m _n ↑∞, n≧1} is a martingale difference array, conditions are given for the distribution and moment convergence of S _n,k to the distribution and moments of where H _k is the Hermite polynomial of degree k and Z is a standard normal variable. This is intimately related to an identity (*) for multiple Wiener integrals. Under alternative conditions, similar results hold for S _{n, k} /U _n ^k and S _{n, k} /V _n ^k where and V _n ² V _n ² is the conditional variance. Research supported by the National Science Foundation under Grant DMS-8601346     

Schauder estimates for sub-elliptic equations

We establish Hölder estimates of second derivatives for a class of sub-elliptic partial differential operators in ${\mathbb{R}^{N}}$ of the kind $$\mathcal L=\sum_{i,j=1}^{m}a_{ij}(x)X_{i}X_{j}+X_{0},$$ where the X _j ’s are smooth vector fields in ${\mathbb{R}^{N}}$ , and a _ij is a uniformly elliptic matrix. It is assumed that the X _j ’s satisfy homogeneity conditions with respect to a group of dilations δ _r which yield the existence of a composition law ${\circ}$ in ${\mathbb{R}^{N}}$ making the triplet ${\mathbb G=(\mathbb{R}^{N},\circ,\delta_{r})}$ an homogeneous Lie group on which the X _j ’s are left translation invariant. The Hölder norms are defined in terms of this composition law. The main tools used are the Taylor formula for smooth functions on ${\mathbb{G}}$ , some properties of the corresponding Taylor polynomials, and an orthogonality theorem that extends to homogeneous Lie groups a classical theorem of Calderón and Zygmund in the Euclidean setting.     

The Law of Iterated Logarithm of Rescaled Range Statistics for AR（1） Model

Let {Xn,n ≥ 0} be an AR（1） process. Let Q（n） be the rescaled range statistic, or the R/S statistic for {Xn} which is given by （max1≤k≤n（∑j=1^k（Xj - ^-Xn）） - min 1≤k≤n（∑j=1^k（ Xj - ^Xn ））） /（n ^-1∑j=1^n（Xj -^-Xn）^2）^1/2 where ^-Xn = n^-1 ∑j=1^nXj. In this paper we show a law of iterated logarithm for rescaled range statistics Q（n） for AR（1） model.     

A class of problems for which cyclic relaxation converges linearly

The method of cyclic relaxation for the minimization of a function depending on several variables cyclically updates the value of each of the variables to its optimum subject to the condition that the remaining variables are fixed. We present a simple and transparent proof for the fact that cyclic relaxation converges linearly to an optimum solution when applied to the minimization of functions of the form for a _i,j ,b _i ,c _i ∈ℝ_≥0 with max {min {b ₁,b ₂,…,b _n },min {c ₁,c ₂,…,c _n }}>0 over the n-dimensional interval [l ₁,u ₁]×[l ₂,u ₂]×⋅⋅⋅×[l _n ,u _n ] with 0<l _i <u _i for 1≤i≤n. Our result generalizes several convergence results that have been observed for algorithms applied to gate- and wire-sizing problems that arise in chip design.     

Circular law theorem for random Markov matrices

Let (X _jk ) _jk≥1 be i.i.d. nonnegative random variables with bounded density, mean m, and finite positive variance σ ². Let M be the n?× n random Markov matrix with i.i.d. rows defined by ${M_{jk}=X_{jk}/(X_{j1}+\cdots+X_{jn})}$ . In particular, when X ₁₁ follows an exponential law, the random matrix M belongs to the Dirichlet Markov Ensemble of random stochastic matrices. Let λ₁, . . . , λ _n be the eigenvalues of ${\sqrt{n}M}$ i.e. the roots in ${\mathbb{C}}$ of its characteristic polynomial. Our main result states that with probability one, the counting probability measure ${\frac{1}{n}\delta_{\lambda_1}+\cdots+\frac{1}{n}\delta_{\lambda_n}}$ converges weakly as n→∞ to the uniform law on the disk ${\{z\in\mathbb{C}:|z|\leq m^{-1}\sigma\}}$ . The bounded density assumption is purely technical and comes from the way we control the operator norm of the resolvent.     

Necessary and sufficient conditions for the asymptotic distribution of the largest entry of a sample correlation matrix

Let {X _k,i ; i ≥ 1, k ≥ 1} be a double array of nondegenerate i.i.d. random variables and let {p _n ; n ≥ 1} be a sequence of positive integers such that n/p _n is bounded away from 0 and ∞. In this paper we give the necessary and sufficient conditions for the asymptotic distribution of the largest entry ${L_{n}={\rm max}_{1\leq i < j\leq p_{n}}|\hat{\rho}^{(n)}_{i,j}|}$ of the sample correlation matrix ${{\bf {\Gamma}}_{n}=(\hat{\rho}_{i,j}^{(n)})_{1\leq i,j\leq p_{n}}}$ where ${\hat{\rho}^{(n)}_{i,j}}$ denotes the Pearson correlation coefficient between (X _1,i , ..., X _n,i )′ and (X _1,j ,...,X _n,j )′. Write ${F(x)= \mathbb{P}(|X_{1,1}|\leq x), x\geq0}$ , ${W_{c,n}={\rm max}_{1\leq i < j\leq p_{n}}|\sum_{k=1}^{n}(X_{k,i}-c)(X_{k,j}-c)|}$ , and ${W_{n}=W_{0,n},n\geq1,c\in(-\infty,\infty)}$ . Under the assumption that ${\mathbb{E}|X_{1,1}|^{2+\delta} < \infty}$ for some δ > 0, we show that the following six statements are equivalent: $$ {\bf (i)} \quad \lim_{n \to \infty} n^{2}\int\limits_{(n \log n)^{1/4}}^{\infty}\left( F^{n-1}(x) - F^{n-1}\left(\frac{\sqrt{n \log n}}{x}\right) \right) dF(x) = 0,$$ $$ {\bf (ii)}\quad n \mathbb{P}\left ( \max_{1 \leq i < j \leq n}|X_{1,i}X_{1,j} | \geq \sqrt{n \log n}\right ) \to 0 \quad{\rm as}\,n \to \infty,$$ $$ {\bf (iii)}\quad \frac{W_{\mu, n}}{\sqrt {n \log n}}\stackrel{\mathbb{P}}{\rightarrow} 2\sigma^{2},$$ $$ {\bf (iv)}\quad \left ( \frac{n}{\log n}\right )^{1/2} L_{n} \stackrel{\mathbb{P}}{\rightarrow} 2,$$ $$ {\bf (v)}\quad \lim_{n \rightarrow \infty}\mathbb{P}\left (\frac{W_{\mu, n}^{2}}{n \sigma^{4}} - a_{n}\leq t \right ) = \exp \left \{ - \frac{1}{\sqrt{8\pi}} e^{-t/2}\right \}, - \infty < t < \infty,$$ $$ {\bf (vi)}\quad \lim_{n \rightarrow \infty}\mathbb{P}\left (n L_{n}^{2} - a_{n}\leq t \right ) = \exp \left \{ - \frac{1}{\sqrt{8 \pi}} e^{-t/2}\right \}, - \infty < t < \infty$$ where ${\mu=\mathbb{E}X_{1,1}, \sigma^{2}=\mathbb{E}(X_{1,1} - \mu)^{2}}$ , and a _n  = 4 log p _n ? log log p _n . The equivalences between (i), (ii), (iii), and (v) assume that only ${\mathbb{E}X_{1,1}^{2} < \infty}$ . Weak laws of large numbers for W _n and L _n , n ≥  1, are also established and these are of the form ${W_{n}/n^{\alpha}\stackrel{\mathbb{P}}{\rightarrow} 0}\,(\alpha > 1/2)$ and ${n^{1-\alpha}L_{n}\stackrel{\mathbb{P}}{\rightarrow} 0}\,(1/2 < \alpha \leq 1)$ , respectively. The current work thus provides weak limit analogues of the strong limit theorems of Li and Rosalsky as well as a necessary and sufficient condition for the asymptotic distribution of L _n obtained by Jiang. Some open problems are also posed.     

A Gaussian Inequality for Expected Absolute Products

We prove the inequality that \mathbbE|X₁X₂?X_n| ￡ ?{per(\varSigma )}{\mathbb{E}}|X_{1}X_{2}\cdots X_{n}|\leq \sqrt{\mathrm{per}(\varSigma )}, for any centered Gaussian random variables X ₁,…,X _n with the covariance matrix Σ, followed by several applications and examples. We also discuss a conjecture on the lower bound of the expectation.     

Weak Invariance Principle for the Local Times of Partial Sums of Markov Chains

Let {X _n } be an integer-valued Markov chain with finite state space. Let $S_{n}=\sum_{k=0}^{n}X_{k}$ and let L _n (x) be the number of times S _k hits x∈? up to step n. Define the normalized local time process l _n (t,x) by The subject of this paper is to prove a functional weak invariance principle for the normalized sequence l _n (t,x), i.e., we prove under the assumption of strong aperiodicity of the Markov chain that the normalized local times converge in distribution to the local time of the Brownian motion.     

A nonclassical LIL for sums of B-valued random variables when extreme terms are excluded

Let {X,X _n ; n≧1} be a sequence of B-valued i.i.d. random variables. Denote $X_{{n}}^{(r)}=X_{{m}}$ if ∥X _m ∥ is the r-th maximum of {∥X _k ∥; k≦n}, and let ${}^{(r)}S_{{n}}=S_{{n}}-(X_{{n}}^{(1)}+\cdots+X_{{n}}^{(r)})$ be the trimmed sums, where $S_{{n}}=\sum_{ k=1}^{n}X_{{k}}$ . Given a sequence of positive constants {h(n), n≧1}, which is monotonically approaching infinity and not asymptotically equivalent to loglogn, a limit result for $^{(r)}S_{{n}}/\sqrt{2nh(n)}$ is derived.     

Tail Behavior of Sums and Maxima of Sums of Dependent Subexponential Random Variables

In this paper, we consider dependent random variables X _k , k=1,2,?? with supports on [?b _k ,??), respectively, where the b _k ??0 are some finite constants. We derive asymptotic results on the tail probabilities of the quantities $S_{n}=\sum_{k=1}^{n} X_{k}$ , X _(n)=max?_1??k??n X _k and S _(n)=max?_1??k??n S _k , n??1 in the case where the random variables are dependent with heavy-tailed (subexponential) distributions, which substantially generalize the results of Ko and Tang (J. Appl. Probab. 45, 85?C94, 2008).     

Large Deviations for Proportions of Observations Which Fall in Random Sets Determined by Order Statistics

Let {X _n :n?≥?1} be independent random variables with common distribution function F and consider $K_{h:n}(D)=\sum_{j=1}^n1_{\{X_j-X_{h:n}\in D\}}$ , where h?∈?{1,...,n}, X _1:k ?≤???≤?X _k:k are the order statistics of the sample X ₁,...,X _k and D is some suitable Borel set of the real line. In this paper we prove that, if F is continuous and strictly increasing in the essential support of the distribution and if $\lim_{n\to\infty}\frac{h_n}{n}=\lambda$ for some λ?∈?[0,1], then $\{K_{h_n:n}(D)/n:n\geq 1\}$ satisfies the large deviation principle. As a by product we derive the large deviation principle for order statistics $\{X_{h_n:n}:n\geq 1\}$ . We also present results for the special case of Bernoulli distributed random variables with mean p?∈?(0,1), and we see that the large deviation principle holds only for p?≥?1/2. We discuss further almost sure convergence of $\{K_{h_n:n}(D)/n:n\geq 1\}$ and some related quantities.     

A Note on Drazin Invertibility for Upper Triangular Block Operators

A bounded linear operator A acting on a Banach space X is said to be an upper triangular block operators of order n, and we write ${A \in \mathcal{UT}_{n}(X)}$ , if there exists a decomposition of ${X = X_{1} \oplus . . . \oplus X_{n}}$ and an n × n matrix operator ${(A_{i,j})_{\rm 1 \leq i, j \leq n}}$ such that ${A = (A_{i, j})_{1 \leq i, j \leq n}, A_{i, j} = 0}$ for i > j. In this note we characterize a large set of entries A _{i, j} with j > i such that ${\sigma_{\rm D} (A) = {\bigcup\limits_{i = 1}^{n}} \sigma_{\rm D} (A_{i, i})}$ ; where σ_D(.) is the Drazin spectrum. Some applications concerning the Fredholm theory and meromorphic operators are given.     

Tame Sets in the Complement of Algebraic Variety

Let A?? ^N be an algebraic variety with dim?A≤N?2. Given discrete sequences {a _j },{b _j }?? ^N \ A with slow growth ( $\sum_{j}{1\over|a_{j}|^{2}}<\infty,\sum_{j}{1\over |b_{j}|^{2}}<\inftyLet A⊂ℂ ^N be an algebraic variety with dim A≤N−2. Given discrete sequences {a _j },{b _j }⊂ℂ ^N \ A with slow growth ( ?_j[1/(|a_j|²)] < ￥,?_j[1/(|b_j|²)] < ￥\sum_{j}{1\over|a_{j}|^{2}}<\infty,\sum_{j}{1\over |b_{j}|^{2}}<\infty ) we construct a holomorphic automorphism F with F(z)=z for all z∈A and F(a _j )=b _j for all j∈ℕ. Additional approximation of a given automorphism on a compact polynomially convex set, fixing A, is also possible. Given unbounded analytic variety A there is a tame set E such that F(E)≠{(j,0 ^N−1):j∈ℕ} for all automorphisms F with F| _A =id. As an application we obtain an embedding of a Stein manifold into the complement of an algebraic variety in ℂ ^N with interpolation on a given discrete set.     

Asymptotic Behavior of the Maximum of Sums of I.I.D. Random Variables along Monotone Blocks

Let {X_i, Y_i}_i=1,2,... be an i.i.d. sequence of bivariate random vectors with P(Y₁ = y) = 0 for all y. Put M_n(j) = max_0≤k≤n-j (X_k+1 + ... X_k+j)I_k,j, where I_k,k+j = I{Y_k+1 < ⋯ < Y_k+j} denotes the indicator function for the event in brackets, 1 ≤ j ≤ n. Let L_n be the largest index l ≤ n for which I_k,k+l = 1 for some k = 0, 1, ..., n - l. The strong law of large numbers for “the maximal gain over the longest increasing runs,” i.e., for M_n(L_n) has been recently derived for the case where X₁ has a finite moment of order 3 + ε, ε > 0. Assuming that X₁ has a finite mean, we prove for any a = 0, 1, ..., that the s.l.l.n. for M(L_n - a) is equivalent to EX ₁ ^3+a I{X₁ > 0} < ∞. We derive also some new results for the a.s. asymptotics of L_n. Bibliography: 5 titles. __________ Translated from Zapiski Nauchnykh Seminarov POMI, Vol. 311, 2004, pp. 179–189.     

Record values and the exponential distribution

A sequence {X _n,n≧1} of independent and identically distributed random variables with continuous cumulative distribution functionF(x) is considered.X _j is a record value of this sequence ifX _j>max (X ₁, …,X _j−1). Let {X _L(n) n≧0} be the sequence of such record values. Some properties ofX _L(n) andX _L(n)−X_L(n−1) are studied when {X _n,n≧1} has the exponential distribution. Characterizations of the exponential distribution are given in terms of the sequence {X _L(n),n≧0} The work was partly completed when the author was at the Department of Statistics, University of Brasilia, Brazil.